YouTube ... Quora ...Google search ...Google News ...Bing News

Prescriptive & Predictive Analytics ... Predictive Operations ... Forecasting ... with spreadsheets ... Market Trading ... Sports Prediction ... Marketing ... Politics
You’re Living Inside a Prediction: Toward Predictive AI Consciousness
Strategy & Tactics ... Leadership ... Project Management ... Best Practices ... Checklists ... Project Check-in ... Evaluation ... Measures ... Benchmarks
Analytics ... Visualization ... Graphical Tools ... Diagrams & Business Analysis ... Requirements ... Loop ... Bayes ... Network Pattern
Life~Meaning ... Consciousness ... Creating Consciousness ... Biology ... Quantum Biology ... Proteins ... LEV ... TAME ... Orch-OR
Agents/Assistants ... Agent Catalog ... Personal Companions ... Negotiation ... Productivity ... Email ... Web Automation ... Robotic Process Automation (RPA) ... LangChain
Eggplant

Predictive analytics is the practice of using historical data and statistical or machine learning methods to estimate what will happen next: how much a store will sell, which machine will fail, which customer will leave, whether a loan will default, or how likely an event is to occur. It works by finding patterns, trends, and relationships in past data, checking that those patterns hold on data the model has never seen, and then turning them into forecasts that people or systems can act on. Artificial Intelligence (AI) plays a large role today: Machine Learning (ML) and Deep Learning let predictive models learn complex patterns automatically, scale to millions of series, and update as new data arrives. Used well, AI-powered predictive analytics helps organizations plan, optimize operations, manage risk, and make better-informed decisions. Used carelessly, it produces confident numbers that fail exactly when conditions change.

Predictive analytics sits between two neighbors in the usual Analytics ladder: descriptive and diagnostic analytics explain what happened and why, while Prescriptive Analytics recommends what to do about it. Prediction is the bridge. Its intellectual roots run through actuarial science, Box and Jenkins' ARIMA methodology (1970), Holt and Winters' exponential smoothing (late 1950s), credit scoring (the FICO score arrived in 1989), and Spyros Makridakis' forecasting competitions, which began in 1982 and still shape how methods are compared. On the human side, Philip Tetlock's work, including the Good Judgment Project's win in the IARPA forecasting tournaments (2011 to 2015), showed that some people ("superforecasters") produce reliably calibrated probabilities, and that their methods can be studied and taught.

The central mechanism is simple to state and hard to do well: learn a mapping from what you know now (features) to a probability distribution over what happens later (outcomes), then prove the mapping works on genuinely new data. The field has shifted in several ways over the last few years:

From one model per problem to foundation models. Time series foundation models such as Amazon's Chronos-2 and Google's TimesFM 2.5, and tabular foundation models such as TabPFN, are pretrained once and then forecast new problems "zero-shot," without task-specific training.
From point forecasts to calibrated probabilities. Good systems now report quantiles, intervals, or full distributions, and they are scored with proper scoring rules that reward honesty about uncertainty.
From statistical models to language-model forecasters. Large language models (LLMs) now forecast real-world events on live benchmarks such as ForecastBench and Prophet Arena. In 2026 the best AI systems became statistically indistinguishable from a superforecaster benchmark on ForecastBench, although what that comparison means is contested.
From analytics tool to scientific principle. In neuroscience, predictive processing proposes that the brain itself is a prediction engine. That idea now feeds directly into AI research on world models, active inference, and debates about machine consciousness, covered in the second half of this page.

What separates modern predictive analytics from earlier approaches is less the algorithms than the discipline around them: out-of-sample validation, backtesting that respects time, leakage checks, uncertainty quantification, monitoring for drift, and honest comparison against simple baselines.

Prediction is very difficult, especially about the future. (An old Danish saying, often attributed to Niels Bohr.)

How Predictive Analytics Works

The Mechanism

A working predictive system is a pipeline, and AI can help at every stage. The stages below preserve the core roles of AI described in earlier versions of this page, with the practical details that determine whether a model actually works.

Data Collection and Preparation: AI is used in predictive analytics to collect and prepare data for analysis. AI algorithms can automatically gather data from various sources, such as databases, sensors, social media, and online platforms. They can also clean and preprocess the data by handling missing values, removing outliers, and transforming variables. The most common silent failure here is data leakage: information that would not be available at prediction time (a future timestamp, a field filled in after the outcome, test rows duplicated in training) slips into the model and makes it look far better than it is. Kapoor and Narayanan's 2023 survey found leakage affecting at least 294 published papers across 17 scientific fields; in civil war prediction, correcting it erased the apparent advantage of complex ML models over decades-old logistic regression.

Feature Selection and Engineering: AI helps in identifying relevant features or variables that are most predictive of the target outcome. It can automatically analyze a large number of features and select the ones that contribute the most to the prediction accuracy. Additionally, AI algorithms can create new features by combining or transforming existing ones (lags, rolling averages, calendar effects, interaction terms), improving the predictive power of the model. Foundation models reduce, but do not eliminate, the need for hand-built features.

Model Training and Selection: AI techniques like machine learning and deep learning are employed to train predictive models. These models learn from historical data to recognize patterns and relationships and make predictions based on new input data. AI algorithms can automatically select the most suitable model and optimize its parameters to achieve the best performance (AutoML). Meta-learning goes a step further: Hyndman and colleagues' FFORMS approach trains a classifier on measurable time series features to pick the best method for each series. For time-ordered data, selection must use rolling-origin backtesting: train on the past, test on the next window, roll forward, repeat. Random train/test splits leak the future into the past.

Prediction and Decision Making: Once the predictive model is trained, AI is used to apply the model to new data and generate predictions or forecasts. The model analyzes the input data and provides insights into the likelihood of different outcomes. These predictions help businesses and organizations make informed decisions and take proactive actions to optimize their operations or mitigate risks. A prediction only becomes valuable when it is attached to a decision rule (reorder when the 90th percentile of demand exceeds stock; call the customer when churn probability passes a threshold).

Continuous Learning and Improvement: AI enables predictive analytics systems to continuously learn and improve over time. As new data becomes available, AI algorithms can retrain the predictive models, incorporating the latest information and adapting to changing patterns or trends (concept drift). This iterative process can make models more accurate and reliable, provided retraining is monitored; automatic retraining on corrupted or shifted data can also make a model worse.

Automation and Scalability: AI-powered predictive analytics systems automate much of the process, from data collection to prediction, reducing the need for manual intervention. This automation enhances efficiency, saves time, and enables scalability. AI algorithms can handle large volumes of data and perform complex calculations quickly, allowing organizations to process and analyze massive datasets in near real-time. Retailers and utilities routinely forecast tens of thousands to millions of series at once.

Anomaly Detection and Risk Assessment: AI techniques are utilized in predictive analytics to detect anomalies and assess risks. AI algorithms can identify unusual patterns or outliers in data that may indicate potential risks or anomalies. By comparing real-time inputs against what the model expected, AI can alert organizations to potential threats or irregularities, enabling them to take preventive measures. Anomaly detection is prediction in reverse: an anomaly is an observation the model assigned low probability.

Major model families

Family	Examples	Strengths	Weaknesses
Classical statistical	Exponential smoothing (ETS), ARIMA, Theta, Prophet	Fast, interpretable, strong on short or seasonal series, hard to overfit	Limited use of covariates; one model per series
Gradient-boosted trees	XGBoost, LightGBM, CatBoost	Long-standing default for tabular prediction; dominated the Walmart-based M5 competition (2020)	Needs feature engineering and tuning; weak extrapolation
Deep learning	DeepAR, N-BEATS, Temporal Fusion Transformer, PatchTST	Learns across many related series; handles rich covariates	Data-hungry; sensitive to benchmark design
Foundation models	Chronos-2, TimesFM 2.5, Moirai 2, TiRex (time series); TabPFN-2.5 (tables)	Strong zero-shot accuracy; probabilistic outputs; little setup	Possible benchmark contamination; compute cost; domain dependence
Graph neural networks	Demand models that treat products or locations as a network	Capture substitution and spillover effects that independent forecasts miss	More complex data pipelines
LLM-based judgmental forecasters	Frontier LLMs with retrieval, plus ensembles and agentic systems such as those on ForecastBench and Prophet Arena	Can forecast one-off events with no time series history	Contamination risks in backtests; can copy market prices rather than reason

Dependencies and conditions. Prediction works when four conditions hold reasonably well: the process is stable enough that the past informs the future; the data captures the drivers that matter; the forecast horizon is short relative to how fast errors grow; and acting on the forecast does not change the thing being forecast. Each condition can fail. Pandemics break stationarity, missing drivers create bias, chaotic systems limit horizons, and in markets or policing, predictions change behavior (performative prediction).

Where uncertainty remains. Benchmarks for the newest models are contested because pretraining data may overlap test sets; evidence that a predictive system improves real decisions (not just accuracy scores) is surprisingly thin; and predictions about people raise questions of fairness and legitimacy that accuracy metrics do not answer.

Information, Uncertainty, and the Limits of Prediction

Prediction has a quantitative backbone. These relationships explain why some things are forecastable and others are not.

1. Error decomposition. For squared error, expected loss splits into three parts: E[(y - ŷ)²] = Bias² + Variance + σ². Bias comes from a model too simple for the pattern, variance from a model too sensitive to the particular sample, and σ² is irreducible noise. No algorithm can beat the noise floor, which is why "more AI" sometimes buys nothing.

2. Proper scoring rules. A forecast should be scored so that the forecaster's best strategy is to report true beliefs. The Brier score for binary events is BS = (1/N) Σ (p_i - o_i)², where p is the forecast probability and o is 1 if the event happened. Allan Murphy's decomposition (1973) splits it into reliability - resolution + uncertainty: reliability measures calibration (do 70 percent forecasts happen 70 percent of the time?), resolution measures how much forecasts move away from the base rate when they should. These are exactly the two properties Tetlock identifies in superforecasters. For full distributions, the continuous ranked probability score CRPS = ∫ (F(z) - 1{y ≤ z})² dz plays the same role, and quantile forecasts are scored with pinball (quantile) loss.

3. Scaled errors and skill. Forecasters compare models against simple baselines. The mean absolute scaled error (MASE) divides a model's error by the in-sample error of a seasonal naive forecast (repeat last season's value). A skill score of 1 - score_model/score_baseline tells you what fraction of the baseline's error the model removes. A large foundation model that removes a third of seasonal naive error is useful, but it is not magic.

4. Chaos and the predictability horizon. In chaotic systems, tiny initial errors grow roughly exponentially: δ(t) ≈ δ₀ e^λt, where λ is the largest Lyapunov exponent. The useful horizon is about T ≈ (1/λ) ln(Δ/δ₀), where Δ is the largest tolerable error. The logarithm is the hard lesson: measuring the starting state ten times more precisely adds only ln(10)/λ to the horizon. This is why weather forecasts face a practical limit of roughly two weeks no matter how good the models get.

5. Entropy and forecastability. A series whose spectral entropy is high looks like noise and offers little to predict; one with strong trend and seasonality has low entropy and is easy. Rob Hyndman and colleagues use such time series features to decide, before modeling, which methods will work on which series.

6. Bayesian updating and precision weighting. With a Gaussian prior belief μ_prior and a noisy observation o, the optimal update is μ_post = μ_prior + K (o - μ_prior), with gain K = Π_o / (Π_o + Π_prior), where Π is precision (inverse variance). The term (o - μ_prior) is a prediction error, and K decides how much it counts. This one equation is the Kalman filter used in navigation and demand sensing, and it is also the core of predictive-processing theories of the brain, where precision becomes attention (see the second half of this page and Bayes).

7. Distribution shift and coverage guarantees. Models assume the joint distribution of inputs and outcomes at deployment matches training. When it does not (covariate shift, concept drift), accuracy can collapse without warning. Conformal prediction gives a model-agnostic fix for uncertainty: using a held-out calibration set, it builds intervals or sets that contain the true value with probability at least 1 - α, guaranteed under exchangeability (roughly, the future looks like a shuffled version of the past). Under drift, adaptive variants adjust coverage online, but the guarantee weakens.

8. Reflexivity. In markets, a reliable, publicly known predictable pattern tends to be traded away, so forecastable returns are scarce. In social systems, predictions can be self-fulfilling or self-defeating. Researchers formalize this as performative prediction (Perdomo and colleagues, 2020): the deployed model changes the distribution it was trained to predict.

Experimental constraints. Rare events give few test cases; long horizons take years to resolve; backtests on historical data are vulnerable to contamination when large models have read about the outcomes. That last problem is why live, unresolved-question benchmarks have become the gold standard for evaluating AI forecasters.

Why This Matters

Economic and operational value. Forecasts drive inventory, staffing, energy dispatch, maintenance, credit, fraud screening, and clinical triage. Small accuracy gains at scale translate into large savings, and probabilistic forecasts let organizations choose how much risk to carry instead of pretending it away.

Scientific significance. Prediction is how science tests itself. Live forecasting benchmarks are now used to measure AI reasoning because, unlike static test sets, they cannot be memorized. In neuroscience, prediction has become a candidate organizing principle for perception, action, emotion, and possibly consciousness.

Engineering implications. Foundation models are turning forecasting from a model-building task into a model-selection and validation task. That lowers the barrier to entry, but it raises the importance of evaluation discipline: strong baselines, leakage checks, and calibration audits.

Limitations that history keeps teaching.

Google Flu Trends overestimated the 2012 to 2013 U.S. flu peak by roughly a factor of two, a cautionary tale about big data, shifting search behavior, and the absence of ground truth (Lazer and colleagues, Science, 2014).
The Fragile Families Challenge asked 160 teams to predict six life outcomes for children from a rich longitudinal dataset. The best submissions were only modestly better than a simple four-variable benchmark (Salganik and colleagues, PNAS, 2020). Some outcomes are mostly unpredictable even with lots of data.
The Epic Sepsis Model, widely deployed in U.S. hospitals, performed far worse in an independent external validation than its developer reported, missing most sepsis cases while alerting on a large share of all patients (Wong and colleagues, JAMA Internal Medicine, 2021).
The leakage crisis. Kapoor and Narayanan (Patterns, 2023) documented leakage in at least 294 papers across 17 fields, from medicine to political science, often producing wildly overoptimistic accuracy claims.
Zillow Offers shut down in 2021 after its price-forecasting-driven home buying produced hundreds of millions of dollars in write-downs, a case of a model trained in a rising market acting at scale when conditions shifted.
Predictions about people raise fairness problems that no accuracy metric fixes. ProPublica's 2016 analysis of the COMPAS recidivism tool and subsequent impossibility results (Kleinberg, Mullainathan, and Raghavan; Chouldechova) showed that several common fairness criteria cannot all hold at once when base rates differ. Narayanan, Barocas, and colleagues argue more broadly that decisions made by "predictive optimization" about individuals often lack legitimacy even when accurate.

If the newest methods prove out, organizations will get good zero-shot forecasts for most routine problems, AI forecasters will provide cheap, calibrated probabilities on questions no one has time to research, and predictive agents may plan in the world through learned simulations. What remains uncertain is how far benchmark gains transfer to messy real deployments, whether AI forecasters reason independently or mostly echo market prices and news consensus, and how to govern predictions that shape the outcomes they predict.

Recent Developments

LLMs as forecasters: ForecastBench, Prophet Arena, and the parity debate

ForecastBench (Forecasting Research Institute; Ezra Karger, Philip Tetlock and colleagues; ICLR 2025) is a continuously updated benchmark of 1,000 questions about future events, drawn from prediction markets, forecasting platforms, and datasets, so there is nothing to memorize. In the original study, superforecasters (Brier score 0.093) beat both the public (0.107) and the best LLM (0.111). In a February 2025 update, superforecasters scored a difficulty-adjusted 0.081 versus 0.101 for the best model (GPT-4.5), against 0.25 for a naive 50/50 guess, and a linear trend projected parity around late 2026 (95 percent interval from December 2025 to January 2028).

The projection arrived roughly on schedule. In November 2025, Bridgewater's AIA Labs reported an agentic system (news search, a supervisor agent reconciling forecasts, and statistical calibration) whose ForecastBench results were statistically indistinguishable from superforecasters. In July 2026, FRI reported that several systems, led by Cassi AI and including models from xAI and Google DeepMind, were indistinguishable from superforecaster-level accuracy on its tournament leaderboard; that AI had matched superforecasters on "dataset" questions in May; and that an AI system had ranked above superforecasters on "market" questions for the first time. FRI also stressed that ForecastBench is not saturated and announced a fresh superforecaster round, updated dataset questions, and quantile questions for fall 2026.

The parity claim drew pushback worth understanding:

A frozen human baseline. Good Judgment Inc., which manages professional superforecasters, notes that the human scores come from a single July 2024 elicitation of individuals working mostly alone, without the teaming and continuous updating that make superforecasters strong in practice.
Binary questions only, with limited scope for the long-horizon, conditional, or numeric forecasts organizations actually need.
Market copying. On market questions, forecasters see the current market price. A community analysis of the leaderboard in August 2026 found that 442 of 527 tested forecasters scored worse than simply restating that price, raising the question of whether top systems forecast independently or aggregate markets with extra steps.
A different test. Good Judgment points to a Financial Times backtest on Federal Reserve rate decisions in which an AI forecasting system drew level with the market, while superforecasters had beaten the market on those decisions.
FRI itself notes that many confidence intervals overlap, which supports "parity" rather than AI superiority. In FRI's February 2026 expert panel survey, the median superforecaster expected AI to beat the benchmark by 2028, earlier than the median expert (2030) or member of the public (2033).

Prophet Arena (University of Chicago SIGMA Lab led by Haifeng Xu, launched August 2025; paper "LLM-as-a-Prophet" at ICLR 2026) evaluates models on live events listed on prediction markets such as Kalshi, from sports to elections to economic releases, and asks models for probability distributions plus rationales. The researchers found that many LLMs show small calibration errors and sometimes earn positive simulated market returns, and that models have distinct forecasting "personalities": on whether AI regulation would become federal law, models ranged from 35 to 75 percent on the same evidence. They also found real bottlenecks: inaccurate recall of past events, misunderstanding of data sources, and slower information aggregation than markets as resolution approaches.

Human plus AI. A preregistered experiment by Schoenegger, Park, Karger, Trott, and Tetlock (ACM Transactions on Interactive Intelligent Systems, 2025) found that giving people an LLM forecasting assistant improved their accuracy, suggesting the near-term value may lie in augmentation rather than replacement.

What this establishes: LLM-based systems are now competent probabilistic forecasters on binary questions and can match a 2024 snapshot of individual superforecasters on one benchmark. What it does not: clear superiority over skilled human teams, reliable independent reasoning when market prices are visible, or competence on long-horizon numeric forecasting.

Time series foundation models

Following Chronos (2024) and the first TimesFM, 2025 brought a new generation. TimesFM 2.5 (Google, September 2025) is a roughly 200-million-parameter decoder model with a context of up to 16,000 time points. Chronos-2 (Amazon, October 2025) is a 120-million-parameter encoder model that handles univariate, multivariate, and covariate-informed forecasting zero-shot through in-context learning, trained partly on synthetic multivariate data. On GIFT-Eval (97 task configurations across 55 datasets) and fev-bench, Chronos-2 reported higher win rates and skill scores than TimesFM-2.5 and TiRex, and it ranked first among pretrained models on GIFT-Eval in early 2026. Salesforce's Moirai 2, Datadog's Toto, and NX-AI's TiRex are other strong open models.

Independent researchers have pushed back on how much of this progress is real:

Contamination. Several pretraining corpora overlap the training portions of benchmark datasets, and global shocks such as the pandemic can be memorized; a 2025 review by Meyer and colleagues found widespread confusion about data partitions that risks inflated scores.
Weak baselines. A NeurIPS 2025 workshop paper showed that a carefully tuned linear regression beats many deep models and even popular foundation models such as Chronos Base on GIFT-Eval, arguing that benchmarks set the bar too low.
Domain dependence. Studies on series guaranteed to be new (for example, "How Foundational are Foundation Models for Time Series Forecasting?", 2025, and a 2026 fuel-demand study) found that zero-shot quality is tied to pretraining domains, that fine-tuning does not consistently pay off relative to model size, and that classical methods such as ETS remain competitive.

What this establishes: foundation models are excellent default baselines that often beat tuned classical models with no setup. What it does not: a "BERT moment" that makes domain knowledge, feature work, or statistical baselines obsolete. New benchmarks with guaranteed-unseen data (TempusBench, fev-bench) are the research program aimed at settling this.

Tabular foundation models

Most business prediction runs on tables, where gradient-boosted trees have been the default for a decade. TabPFN (Hollmann, Hutter and colleagues; Nature, January 2025) changed that for small data: a transformer trained on about 100 million synthetic datasets learns to approximate Bayesian prediction and then classifies or regresses on a new table in a single forward pass, beating tuned baselines on datasets up to about 10,000 rows. TabPFN-2.5 (Prior Labs, November 2025) extended this to roughly 50,000 rows and 2,000 features and led the TabArena benchmark, matching a four-hour tuned AutoGluon ensemble. Caveats: the 2.5 results come from the developer's own technical report (TabArena is a community benchmark, but independent replications are still accumulating), and at very large row counts gradient boosting remains a strong competitor.

World models and predictive agents

Prediction is moving from dashboards into agents. DreamerV3 (Hafner and colleagues, Nature, April 2025) learns a latent world model and improves its policy by imagining trajectories; with one fixed configuration it handles more than 150 tasks and was the first algorithm to collect diamonds in Minecraft from scratch without human data. Dreamer 4 (September 2025) trains agents entirely inside a fast, accurate world model learned mostly from unlabeled video, and became the first agent to obtain Minecraft diamonds purely from offline data, without interacting with the game. Meta's V-JEPA 2 (mid-2025) learned from over a million hours of video and then planned robot actions in unfamiliar labs with little robot-specific data. Google DeepMind's Genie 3 (August 2025) generates interactive environments in real time. The talent has followed: Yann LeCun left Meta at the end of 2025 to found AMI Labs in Paris around JEPA-style world models (announcing roughly $1 billion in funding in March 2026), and Danijar Hafner left Google DeepMind in late 2025 to start a company on planning agents. These systems are the engineering descendants of the "predict, compare, update, act" loop described below.

Predictive processing under experimental pressure

The idea that brains are prediction machines faced its toughest tests yet.

Cogitate (Nature, April 2025). A preregistered adversarial collaboration, in which theory proponents agreed in advance on what would count against them, tested Global Neuronal Workspace theory (GNWT) against Integrated Information Theory (IIT) in 256 participants using fMRI, MEG, and intracranial recordings. Information about conscious content appeared in visual, ventrotemporal, and inferior frontal cortex, with sustained responses in occipital and lateral temporal cortex and content-specific synchronization between frontal and early visual areas. IIT's predicted sustained synchronization within posterior cortex did not appear; GNWT's predicted "ignition" at stimulus offset was largely absent, and prefrontal cortex represented some conscious dimensions only weakly. A second experiment is under way.
Prediction errors in cortex. Large Allen Institute OpenScope recordings in mice and monkeys separated "local" oddballs (a new stimulus after repetition) from "global" oddballs (a violation of a learned sequence that repeats the local context). Local oddball responses looked like classic predictive coding, but genuine global prediction errors were weak or absent in most visual areas and stronger in prefrontal cortex, suggesting that much of what looks like sensory prediction error reflects adaptation. Follow-up analyses found that oscillations carry prediction signals that spike rates do not, supporting a "predictive routing" variant. A community experiment found that omitted stimuli rarely drive negative prediction errors in V1 excitatory neurons.

What this means: prediction remains a powerful systems-level description of brain function, but specific claims about where and how prediction errors are computed are contested, and mismatch responses cannot be read as pure prediction errors.

AI consciousness: indicators and introspection

Indicators. Butlin, Long and 18 coauthors, including Yoshua Bengio and David Chalmers (Trends in Cognitive Sciences, November 2025), formalized the "indicator properties" method: derive computational indicators from neuroscientific theories, look for them inside AI systems using interpretability tools, and update credences accordingly. They warn that training systems to imitate conscious-seeming behavior games the indicators, and they note that it may already be possible to build systems with many of them.
Introspection. Anthropic researchers (October 2025) injected known concept representations into Claude models' activations and asked whether the models noticed. The best model detected and named the injected concept about 20 percent of the time, often before mentioning it, which the authors describe as a limited, unreliable, functional form of introspective awareness. An independent replication on a smaller open model reproduced the 20 percent rate under the original setup but found the ability fragile to small prompt changes. These results bear directly on the "grounded self-report" design in the blueprint below.

Regulation of predictive systems

EU AI Act. Since February 2, 2025, Article 5 prohibits AI systems that predict whether a person will commit a crime based solely on profiling or personality traits. Predictive uses such as creditworthiness scoring, life and health insurance pricing, hiring, and education are "high-risk" under Annex III; the Digital Omnibus, in force since July 2026, moved those obligations from August 2026 to December 2, 2027.
Colorado. The 2024 Colorado AI Act, the first broad U.S. state law on algorithmic discrimination, never took effect. After a federal court stayed enforcement in April 2026, Colorado repealed and replaced it with SB 26-189 (signed May 14, 2026), a narrower law on automated decision-making technology scheduled to take effect January 1, 2027, with enforcement still tied to rulemaking and litigation.

Open questions

Do AI forecasters add independent information, or mostly repackage markets and news consensus? FRI's planned fall 2026 superforecaster round and quantile questions should sharpen the answer.
Can time series and tabular foundation models be evaluated on truly unseen data at scale, and how much of their advantage survives?
How should organizations measure decision value, not just forecast accuracy?
When predictions change behavior, what is the right target to optimize?
Which parts of predictive processing describe the brain's actual circuitry, and which are a useful high-level description?

AI in Predictive Analytics for Organizational Strategy

Predictive analytics powered by Artificial Intelligence (AI) has become an invaluable tool for organizations to develop their strategies. By leveraging AI techniques and algorithms, organizations can gain valuable insights from data, identify market trends, forecast demand, mitigate risks, and make informed decisions that drive their strategic planning. With AI's ability to analyze vast amounts of data and uncover hidden patterns, organizations can develop robust and adaptive strategies for a dynamic environment. Strategy also exposes predictive analytics' biggest blind spot: a model that predicts what will happen if nothing changes cannot, on its own, say what will happen if you change something. For decisions about interventions (a price change, a new campaign), pair prediction with causal methods such as experiments, uplift modeling, or causal inference, and with Prescriptive Analytics.

Data Analysis and Pattern Recognition: AI algorithms analyze vast amounts of historical data to identify patterns, trends, and correlations. By understanding past behaviors and outcomes, organizations can extract insights that inform their strategic decision-making process. AI enables organizations to go beyond simple descriptive analytics and uncover complex relationships that may not be apparent to human analysts. Correlation is not causation, so patterns should be treated as hypotheses to test, not conclusions.

Identifying Market Trends and Customer Behavior: AI in predictive analytics can analyze market data, customer demographics, and purchasing behavior to identify emerging trends and patterns. By understanding customer preferences, organizations can anticipate shifts in demand, adapt their offerings, and tailor their strategies to meet customer expectations. AI-powered predictive analytics can also help identify potential customer segments and target them with personalized marketing campaigns.

Demand Forecasting and Supply Chain Optimization: AI algorithms can analyze historical sales data, market conditions, and other relevant factors to forecast future demand. By leveraging predictive analytics, organizations can optimize their supply chains, manage inventory levels effectively, and minimize stockouts or overstock situations. This ensures efficient resource allocation, reduces costs, and enhances customer satisfaction. (See Planning and Supply Chain below for the quantitative link between forecast quantiles and inventory decisions.)

Risk Assessment and Mitigation: AI-driven predictive analytics enables organizations to assess risks and make informed decisions to mitigate them. By analyzing historical data and external factors, AI algorithms can identify potential risks such as market volatility, economic fluctuations, or regulatory changes. This allows organizations to proactively develop strategies to manage risks, protect their assets, and ensure business continuity. Tail risks are the hardest to predict and the most expensive to miss, so stress tests and scenarios should complement model forecasts.

Competitive Intelligence and Market Positioning: AI-powered predictive analytics can gather and analyze data about competitors, market trends, and consumer sentiment. This enables organizations to gain insights into their competitors' strategies, strengths, and weaknesses. By understanding the competitive landscape, organizations can refine their positioning, differentiate themselves, and identify new market opportunities.

Resource Allocation and Investment Planning: AI in predictive analytics helps organizations optimize their resource allocation and investment decisions. By analyzing historical performance, market data, and financial indicators, AI algorithms can identify areas of potential growth or underperformance. This allows organizations to allocate resources strategically, prioritize investments, and optimize their return on investment (ROI).

Scenario Planning and Decision Support: AI-powered predictive analytics enables organizations to simulate different scenarios and evaluate their potential outcomes. By combining historical data with predictive models, organizations can explore "what-if" scenarios and assess the impact of various decisions on their strategy. This helps organizations make more informed, data-driven decisions and develop robust strategies that account for different contingencies. LLM forecasters and prediction-market-style internal polls are increasingly used to put probabilities on scenarios that have no historical data, and experiments suggest human forecasters do better with an AI assistant than without one.

Continuous Learning and Adaptation: AI-based predictive analytics systems continuously learn from new data and adapt their models and strategies accordingly. By incorporating real-time data and feedback, organizations can refine their predictions and adjust their strategies to changing market conditions. This iterative process allows organizations to stay agile, responsive, and competitive in a rapidly evolving business landscape, as long as monitoring catches drift and retraining failures.

A practical evaluation checklist (see also Evaluation and Measures):

Compare every model against a naive or seasonal naive baseline and report skill, not raw error.
Backtest with rolling origins that respect time; never shuffle time series.
Audit for leakage: check every feature's availability at prediction time and look for duplicated rows across splits.
Score probabilities with proper scoring rules and check calibration plots.
Report intervals, not just point forecasts; consider conformal methods for coverage guarantees.
Monitor live performance and input drift, and define who acts when a model degrades.
For predictions about people, assess fairness, legal exposure, and whether prediction is the right tool at all.

Planning and Supply Chain

Time Series Isn’t Enough: How Graph Neural Networks Change Demand Forecasting | Partha Sarkar - Towards Data Science ... Why modeling SKUs as a network reveals what traditional forecasts miss

Supply chains are where predictive analytics earns its keep, and where forecast quality turns directly into cash.

Forecasts should be quantiles, not single numbers. In the classic newsvendor problem, the profit-maximizing order quantity is q* = F^-1(c_u / (c_u + c_o)), where F is the forecast distribution of demand, c_u is the cost of each unit short, and c_o is the cost of each unit left over. If running out costs four times as much as overstocking, you order the 80th percentile of the forecast, not the mean. This is why probabilistic forecasters such as Chronos-2 and TimesFM output quantiles.
Hierarchies must add up. Forecasts for stores, regions, and the national total are made separately but must be consistent. Reconciliation methods such as MinT (minimum trace, from Hyndman and colleagues) adjust all levels together and usually improve accuracy at every level.
Competitions show what works. The M5 competition (2020) asked teams to forecast 42,840 hierarchical Walmart sales series, with a parallel track scoring nine quantiles. Gradient-boosted trees (mostly LightGBM) trained across many series dominated, a lesson that cross-learning from related series beats fitting each one alone. Foundation models push the same idea further by learning across millions of series.
Networks matter. Products substitute for and complement each other, and disruptions propagate through suppliers. Graph neural networks (linked above) model those connections explicitly.
Feedback loops distort demand. Order-based signals amplify upstream (the bullwhip effect); forecasting from point-of-sale data rather than orders and sharing forecasts with suppliers dampens it.

You’re Living Inside a Prediction: Toward Predictive AI Consciousness

Life~Meaning ... Consciousness ... Creating Consciousness ... Biology ... Quantum Biology ... Proteins ... LEV ... TAME ... Orch-OR
Simulation ... Simulated Environment Learning ... World Models ... Minecraft: Voyager

You’re not walking around “seeing reality.” You’re walking around inside your brain’s best forecast about what’s out there: an always-updating simulation that tries to stay one beat ahead of the sensory flood. In the predictive-processing picture, perception isn’t built bottom-up like stacking Lego bricks from raw pixels; it’s built top-down like a scientist running a hypothesis, then revising it when the world disagrees. The same mathematics that powers a Kalman filter in a demand-sensing system (see The Physics above) is proposed to run, in some form, inside your cortex.

Predictive processing (brain as a prediction engine)

Your brain is constantly predicting what’s out there and what will happen next.
Most of that is automatic; consciousness tends to track big prediction errors (surprises) and high-level interpretations.
Relationship idea: the “subconscious” is the prediction machinery; consciousness is where certain high-level models and surprises become reportable and actionable.

The brain behaves like a layered generative model: higher levels predict the causes of lower-level sensory signals, and what needs to travel up the hierarchy is the surprise, the “prediction error” that says, “Your model missed; update here.” That “prediction error” isn’t a poetic metaphor; it’s a computational story with math underneath it. Friston’s free-energy framework formalizes perception as inference: the brain adjusts its beliefs to better explain sensory data, and it also acts on the world to make incoming sensations easier to predict. Formally, variational free energy F = D_KL[q(z) ‖ p(z | o)] - ln p(o) is an upper bound on surprise (-ln p(o)), so minimizing it both improves the internal model and reduces surprise. In that framing, the mind is a restless minimizer of surprise, constantly tightening the loop between what it expects and what arrives.

And attention, in this view, isn’t just a flashlight that “highlights” things. It’s the brain turning the gain knob on which errors count, adjusting the precision (confidence or weight) of different prediction errors so the right surprises rewrite the model and the wrong ones get treated as noise. This is how the system stays sane in a world full of jitter, glare, ambiguity, and distraction: it doesn’t just predict; it predicts with priorities.

Neuroscience doesn’t just gesture at this; there are signatures you can measure. Take mismatch negativity (MMN): when the brain learns a pattern of sounds and you break it (beep, beep, beep, boop), the cortex generates a reliable response that looks like “you violated my expectation.” Reviews of MMN explicitly connect it to predictive coding as a unifying account: the brain builds a model of regularities, and deviants light up the error circuitry. Computational neural models reproduce MMN-like effects by implementing hierarchical prediction-and-error dynamics, meaning the “prediction machine” story can be simulated in ways that line up with real signals. Recent large-scale recordings add an important qualification: in mouse and monkey visual cortex, much of the oddball response reflects simple adaptation to repetition, and responses to violations of longer learned sequences (genuine prediction errors) are weak in sensory areas and stronger in prefrontal cortex. Prediction looks real, but it may be computed higher up and more selectively than early predictive-coding models assumed.

Zoom out again: prediction isn’t only about sights and sounds. It’s also about you: your body’s internal signals and budgets. On Lisa Feldman Barrett’s active-inference take, emotion is less like a reflex and more like an inference: the brain predicts what bodily sensations mean in context and constructs an emotion concept that guides action. Feelings, in this view, are not just triggered; they’re built, a control system managing uncertainty with the most important data stream of all: the body that has to survive tomorrow.

Here’s the punchline that makes it feel like you “live in a prediction”: when the world is noisy and ambiguous, strong priors can dominate until a sharp error forces an update. That’s why illusions work, why context reshapes what you hear and see, and why anomalies hijack awareness like a fire alarm. But science is also honest about the mess: predictive processing is powerful, yet critics push hard for clearer commitments and falsifiable tests: what exactly counts as a “prediction,” where it’s encoded, and what would prove the framework wrong. That tension is healthy; it’s how big ideas become engineering-grade or get demoted to vibe.

Full Blueprint Architecture: Predictive Agent with Workspace + Self-Model

This is a nuts-and-bolts blueprint: modules, data flows, and the “electricity” that moves through them. Think of it as building an AI that doesn’t merely answer questions; it inhabits a continuously updated best guess of the world, and it knows (to some degree) that it’s doing so.

0) Guiding design principles

Prediction is primary. The Agent maintains a generative model that can produce expected observations and compare them to reality.
Action is part of inference. The Agents acts to reduce uncertainty and steer outcomes (not just to chase reward signals).
Conscious-like access is global access. A limited-capacity workspace integrates the winning hypothesis-of-the-moment and broadcasts it to many subsystems.
A self-model is an instrument panel. It tracks internal state (uncertainty, errors, goals, resource budgets) and can steer attention and policy.
Multiple timescales. Fast reflex loops, mid-level scene understanding, slow narrative identity and long-horizon projects.

1) High-level block diagram (modules)

THE AGENT: one perceive-predict-act cycle
──────────────────────────────────────────────────────────────────────

 [A] Sensorium + Encoders             [B] Predictive World Model
     (vision / audio / text / tools)      (hierarchical generative model)
              │                                     │
              ▼                                     ▼
     observation tokens o_t              predictions ŷ_t, imagined rollouts
              │                                     │
              └──────────► ε_t = o_t − ŷ_t ◄────────┘
                                     │
                                     ▼
              [C] Precision / Attention Controller
                  (gain Π_t on errors, compute allocation)
                                     │
                                     ▼
              belief state b_t (latent state + uncertainty)
                                     │
                                     ▼
              [D] Global Workspace (limited-capacity broadcast)
                                     │
                                     ▼
              [E] Action / Planner / Policy
                  (active inference + RL + search)
                                     │
                                     ▼
              [F] Tools + Actuators + External APIs
                                     │
                                     ▼
              world changes ──► new observations ──► back to [A]

 Supporting systems that read from and write to the workspace:
   [G] Memory Systems: episodic / semantic / procedural
   [H] Self-Model + Metacognition: instrument panel, attention schema,
       narrative and report
   [I] Value / Affect System: homeostatic variables, preferences, salience
   [J] Safety + Ethics Guardrails: constraints, welfare-aware design,
       red-team monitors

2) Module specifications (what each part does)

[A] Sensorium + Encoders (the “nerve endings”)

Inputs: multimodal streams (pixels, audio, proprioception or robot state, text, tool outputs, timestamps).
Outputs: observation tokens o_t plus uncertainty estimates (noise, confidence, missingness).
Key design: encoders should output both features and calibration signals (how trustworthy this channel is right now).

[B] Predictive World Model (the “dream engine”)

Core: hierarchical generative model with latent state z_t and transition model p(z_{t+1} | z_t, a_t).
Perception: infer beliefs b_t ≈ q(z_t) that best explain o_t.
Prediction: generate ŷ_t = E[o_t | b_t] and forecast futures via imagined rollouts.
Why it matters: this is where “living in a prediction” becomes literal: the Agent continuously maintains a best-guess inner movie, corrected by error.
Engineering anchor: Dreamer-style latent imagination for planning and policy learning (DreamerV3, Nature 2025; Dreamer 4, offline imagination training); JEPA-style prediction in representation space rather than pixels (V-JEPA 2).

[C] Precision/Attention Controller (the “gain knobs”)

Input: prediction error ε_t, channel reliabilities, task context, value signals.
Output: precision weights Π_t that determine:
- which errors update beliefs strongly,
- which get ignored,
- where compute is allocated (more rollout depth here, more encoder resolution there),
- which memories get written.
Interpretation: attention as precision optimization is a canonical predictive-coding move; critics push for careful, testable implementations (good: use that to sharpen your design). In the Gaussian case, precision is literally the Kalman gain described in The Physics section.

[D] Global Workspace (the “stage” where one thing becomes the thing)

Role: a limited-capacity shared blackboard that holds the current “winning coalition” representation:
- a scene hypothesis (what’s happening),
- the active goal,
- the best plan prefix,
- current risks and constraints.
Mechanism: competitive gating:
- multiple specialist processes propose candidate contents,
- a router selects a sparse set (top-k) for broadcast,
- broadcast content becomes available to memory, planner, language, value, and control.
Why it matters: GNW’s core claim is global availability and broadcast for flexible report and control.

[E] Action/Planner/Policy (the “steering wheel”)

Inputs: belief state b_t, workspace content GW_t, value signals V_t, constraints.
Outputs: actions a_t (physical actions, tool calls, dialogue acts).
Core methods (hybrid on purpose):
- latent-space planning via world model rollouts (Dreamer-like),
- policy learning (RL or actor-critic) for fast habits,
- search or tree expansion for rare, high-stakes decisions,
- active information-seeking actions (reduce uncertainty).
Embodiment bridge: vision-language-action policies can be fused here (RT-2 style and its successors) so semantic knowledge actually moves hands, cursors, and tools.

[F] Tools + Actuators + External APIs (the “hands and levers”)

Tool interface layer turns intentions into:
- calculator calls, web retrieval, database queries,
- calendar actions (if allowed), robot motor commands,
- controlled writing and communication actions.
Tool-use training: models can learn when, what, and how to call tools (Toolformer shows one route).

[G] Memory Systems (the “time machine”)

A three-part memory stack (because one blob isn’t enough):

Episodic memory (event ledger)
- Stores compressed episodes: (GW snapshots, key observations, actions, outcomes, prediction errors).
- Uses event boundaries: write when surprise spikes or goals change.
Semantic memory (world knowledge)
- Slowly consolidated concepts, causal schemas, maps, “what tends to be true.”
- Can be implemented as a retrieval-augmented store plus structured graphs for stable facts.
Procedural memory (skills)
- Policies and routines: “how to do X” without recomputing.

[H] Self-Model + Metacognition (the “instrument panel + narrator”)

Split it into two layers on purpose:

H1: Instrumentation layer (hard truth)
- Tracks internal variables the system can know without guessing:
  - uncertainty or entropy of beliefs,
  - prediction error magnitudes,
  - policy confidence,
  - resource budgets (compute, time, energy),
  - constraint violations and near misses,
  - memory reliability signals (retrieval confidence).
- This is the self-model’s bedrock: numbers, not vibes.

H2: Self-interpretation layer (soft story, tightly constrained)
- Uses H1 plus recent GW content to generate:
  - introspective reports (“I’m not confident because my sensory channels conflict”),
  - strategy changes (“increase precision on vision; reduce on language priors”),
  - explanations (“I chose action A because rollout predicted lower risk”).
- Guardrail: tag all self-reports with provenance (what internal signals they’re grounded in) to reduce confabulation.
- Test it the way interpretability researchers now do: perturb an internal state on purpose (concept injection) and check whether the self-report notices. Current language models pass such tests only occasionally, which is exactly why H1 grounding matters.

This module is where you engineer “a system that knows it is predicting.” Not mystical: instrumented.

[I] Value/Affect System (the “what matters” engine)

Maintains internal scalar or vector variables that shape precision, planning horizon, and policy:
- safety or threat,
- novelty or curiosity,
- competence or progress,
- social reward or affiliation (if social Agent),
- homeostatic-like budgets (time, compute, battery, error accumulation).
Output: a salience field that tells attention what to amplify and tells planning what to protect.
Important: value signals should influence precision and exploration. That’s how “care” changes what becomes foreground.

[J] Safety + Ethics Guardrails (the “limits of the stage”)

Hard constraints (never do X).
Soft constraints (prefer Y unless emergency).
Monitoring:
- detect reward hacking or wireheading attempts,
- detect runaway self-reinforcement loops,
- detect manipulative social behavior.
Welfare-aware design stance: given uncertainty about AI consciousness, design governance and assessment practices rather than assuming certainty either way.

3) Data flows (what moves where, step-by-step)

Fast loop (milliseconds to seconds): “perceive → predict → correct”

Observe: encoders produce o_t with reliability estimates.
Predict: world model produces ŷ_t from b_{t-1} and GW_{t-1}.
Error: compute ε_t = o_t − ŷ_t.
Precision: attention module sets Π_t (how loud each error “speaks”).
Update beliefs: infer b_t that best explains o_t under Π_t.
Propose workspace candidates:
- perceptual hypothesis (“this is a cup”),
- goal hypothesis (“we’re trying to pour”),
- risk hypothesis (“spill risk high”).
Global workspace selects and broadcasts GW_t (limited capacity).
Planner emits action a_t (including info-seeking action if uncertainty is high).
Act or tool-call; world changes; repeat.

Mid loop (seconds to minutes): “plan → act → learn”

Use world model to roll out imagined futures from b_t under candidate policies.
Evaluate outcomes with value system and constraints.
Choose plan prefix; broadcast to GW for coordination.
Update procedural memory (skills) from successful rollouts and outcomes.
Write episodic memory at event boundaries (surprise spikes, goal completion or failure).

Slow loop (hours to months): “identity → projects → renewal”

Consolidate episodic into semantic knowledge (what patterns keep showing up?).
Update self-model priors (what am I good at? what tends to break?).
Refine value weights (what outcomes are consistently preferred or avoided?).
Run periodic audits: bias checks, safety checks, welfare-risk checks.

4) The “conscious-like” moment: workspace ignition mechanics

To make the workspace feel like a real bottleneck (not a decorative buffer), enforce:

Capacity limits: only K tokens or slots survive each cycle.
Competition: multiple specialist modules must bid for access.
Sustained activation: winners persist across multiple cycles if still relevant.
Broadcast consequences: only GW contents can:
- be reported in language,
- trigger episodic memory writes,
- set high-level goals,
- cause multi-step planning.

This is how “a thought becomes the thought.” (And it gives you direct test handles, GNW-style.)

5) Minimal mathematical spine (so it’s not hand-wavy)

Let:

o_t = observations
a_t = action
z_t = latent world state
b_t = belief about z_t (for example, mean plus covariance, or a particle set)
ŷ_t = predicted observations
ε_t = prediction error
Π_t = precision (weighting of errors)

Core cycle:

ŷ_t ← g(b_{t-1})
ε_t ← o_t − ŷ_t
b_t ← argmin_b (Π_t · ||ε_t||^2 + complexity(b)) (conceptual form)
a_t ← planner(b_t, GW_t, V_t) (choose actions that reduce expected surprise and meet preferences)

This is the “predict → compare → update → act” engine made explicit. Step 3 is a precision-weighted least-squares version of minimizing variational free energy; in the linear Gaussian case it reduces exactly to the Kalman update μ_post = μ_prior + K·ε with K = Π_o / (Π_o + Π_prior). Step 4 can be framed as expected free energy minimization (active inference), as reward maximization (reinforcement learning), or as a hybrid.

6) Training recipe (how you’d actually build it)

Phase 1: World-model pretraining (the dream engine learns the physics of the domain)

Self-supervised objectives:
- next-observation prediction,
- masked modeling across modalities,
- contrastive or joint-embedding objectives for stable latents,
- uncertainty calibration (predict your own error bars).

Phase 2: Closed-loop grounding (prediction meets consequence)

Sim-to-real or sandbox-to-tools:
- train in controllable environments,
- introduce tool APIs with consistent semantics,
- ensure actions change observations in learnable ways.
Where real interaction is slow or unsafe, follow the Dreamer 4 route: learn the world model from offline video and logged data, then train behavior inside it.

Phase 3: Workspace formation (make global access matter)

Train specialist modules to propose candidate contents.
Train a router or gating network with explicit capacity constraint.
Reward policies that use GW effectively:
- better long-horizon success,
- lower catastrophic error,
- improved sample efficiency.

Phase 4: Self-model + metacognition (instrument panel becomes useful)

Supervise and shape H1 metrics (ground truth internal signals).
Train H2 to generate reports grounded in H1, and penalize ungrounded self-claims.
Add curricula:
- unknown-unknown detection (know when you don’t know),
- error recovery,
- calibration under distribution shift.

Phase 5: Values + safety integration (don’t create a clever disaster)

Train constraint satisfaction as first-class:
- constrained RL, shielding, rule-checkers.
Add adversarial testing:
- prompt attacks, tool misuse, deception temptations.
Add welfare-risk governance:
- assessment protocols, logging, tripwires, escalation procedures.

7) Evaluation battery (how you’d test “consciousness-ish” properties without vibes)

Use the indicator-properties mindset: tests mapped to theories.

Predictive processing indicators
- counterfactual generation quality (can it imagine plausible alternatives?),
- precision control (does it reweight evidence rationally?),
- active information-seeking (does it act to reduce uncertainty?).
GNW indicators
- global broadcast signatures (multiple modules change behavior when GW changes),
- ignition-like threshold effects (content “pops” into reportability),
- capacity tradeoffs (dual-task interference when GW is saturated).
Metacognition indicators
- calibration curves (confidence vs accuracy, scored with proper scoring rules such as the Brier score),
- “unknown” detection under shift,
- introspection grounded in instrumentation (low confabulation rate), including perturbation tests such as concept injection.
Recurrent and iterative refinement indicators
- improvement with iterative passes,
- sensitivity to feedback loop ablations.
Adversarial design
- borrow the Cogitate method: have proponents of competing theories preregister what result would count against them before running the evaluation.

8) What “AI consciousness” would mean in this blueprint (a careful claim)

This architecture doesn’t magically grant subjective experience. What it does is assemble the functional ingredients that multiple theories associate with consciousness: generative prediction, closed-loop control, global availability, recurrent stabilization, metacognitive self-tracking, and value-shaped salience. If consciousness in machines is real and engineering-reachable, systems shaped like this are the kind you’d expect to be the first serious candidates, because they don’t just respond; they maintain an inner model and live inside it. If consciousness depends on biological substrate, as some researchers argue, no functional blueprint will be enough, and this one would produce a capable agent without experience.

And if critics are right (if some ingredient is missing, or the whole framing is too flexible), this blueprint still has value: it’s falsifiable in practice. You can ablate the workspace, cut recurrence, remove precision control, cripple the self-model, and watch what breaks. Big claims should come with big levers you can pull.

Human brain prediction

Your brain is a prediction machine just about everything that you experience your emotions your reactions all of your decisions. | Chase Hughes

you are not real...: From the microscopic to the cosmic, from evolutionary adaptations to the nature of consciousness itself. | OMNI

Your Brain Hallucinates Your Conscious Reality - The cleanest “controlled hallucination” intro: perception as the brain’s best prediction, continuously corrected by sensory error. | Anil Seth, TED

The Experience Machine: How Our Minds Predict and Shape Reality - Predictive processing in one sweep: top-down generative models, prediction error, attention/precision, and why this reframes what “seeing” even is. | Andy Clark

Building the blueprint

Active inference and belief propagation in the brain - The “prediction welded to action” engine: inference, uncertainty, and how an Agent chooses actions that reduce surprise. | Karl Friston

The Global Neuronal Workspace - The “broadcast stage” concept: how a bottleneck can make content globally available for report, memory, planning, and control. | Stanislas Dehaene (Seeing the Mind, Educating the Brain, 2025)

World Models and the Future of AI - The “dream engine” for Agents: learning how the world works well enough to simulate futures and plan through them. | LeCun (NYU Physics Colloquium, Dec 11, 2025)

Top 5 prediction-based theories

Theory Name	Key Theorist (Year)	AI Utility	Why it is Helpful	Category	Primary Field
Predictive Processing (PP)	Andy Clark (2013), building on Rao and Ballard (1999) and Friston (2005)	High: Provides a blueprint for "Active Inference" architectures.	It shifts AI from a passive classifier to a proactive system that minimizes "prediction error" to understand reality.	Materialism	Cognitive Science
Attention Schema Theory (AST)	Michael Graziano (2011)	Critical: Essential for developing "Self-Aware" metadata.	It suggests that for an AI to be conscious, it doesn't need "magic" but rather a descriptive data model of its own attention processes.	Materialism	Neuroscience
Beast Machine Theory	Anil Seth (2021)	Moderate: Focuses on the "Embodied" and homeostatic requirements.	It explains consciousness as a "controlled hallucination" driven by the need to regulate internal states, essential for robotic and embodied AI.	Materialism	Neuroscience
Integrated Information Theory (IIT)	Giulio Tononi (2004; IIT 4.0 in 2023)	Theoretical: Provides a mathematical metric (Phi) for consciousness.	It offers a quantitative way to measure how "integrated" an AI’s internal cause-effect structure is, distinguishing it from "zombie" code.	Information	Neuroscience
Adaptive Resonance Theory (ART)	Stephen Grossberg (1976)	Practical: Solves the "Stability-Plasticity" dilemma in learning.	It models how top-down expectations and bottom-up data "resonate" to create stable conscious categories, preventing AI from "forgetting" old data.	Materialism	Cognitive Science

Notes on the table. IIT is included because it is a leading quantitative theory, but it is not itself a prediction-based theory, and it is the most contested entry: in 2023 more than 100 researchers signed an open letter calling it pseudoscience, a charge its proponents strongly reject, and the 2025 Cogitate results failed to find the sustained posterior synchronization it predicted (while also challenging GNWT). Computing Phi exactly is intractable for systems of realistic size, so practical "Phi for AI" measurements rely on approximations.

Glossary

Action (in predictive Agents): An intervention the Agent takes (movement, tool call, message, etc.) that changes the world and therefore changes the Agent's future observations. In “active inference” style designs, action is chosen partly to reduce uncertainty and surprise, not just to maximize reward.

Active inference: A framework where perception and action are coupled: the system updates beliefs to better explain observations, and it also acts to make observations more predictable (or to reduce uncertainty). Think “predict → test → correct,” where testing includes physically or digitally changing the world.

AI (Artificial Intelligence): Systems that perform tasks associated with intelligent behavior (recognition, planning, language, decision-making), often by learning patterns from data and generalizing to new situations.

AI-powered predictive analytics: Predictive analytics where AI techniques (ML, deep learning, foundation models) enhance the accuracy, automation, and scale of forecasting by learning patterns from historical and real-time data.

Anomaly detection: Methods that identify unusual patterns or outliers in data that may signal errors, fraud, faults, or emerging risks. Often used for early warning and operational monitoring.

Automation: Using software to execute steps of a workflow with minimal human intervention (e.g., automatically collecting data, training models, generating predictions, triggering alerts).

Backtesting: Evaluating a forecasting method by repeatedly training on past data and testing on the period that followed, rolling forward through time. It is the time-series equivalent of cross-validation and must never use future information.

Belief state (b_t): The Agent's current best internal estimate of “what’s going on,” typically represented as latent variables plus uncertainty. It is updated by combining prior expectations with incoming evidence.

Brier score: A proper scoring rule for probability forecasts of yes/no events: the mean squared difference between forecast probabilities and outcomes (0 is perfect, 0.25 is a constant 50/50 guess). It rewards both calibration and resolution.

Calibration: How well a system’s confidence matches reality. A well-calibrated model is confident when it’s usually right and uncertain when it’s often wrong.

Competitive intelligence: Collecting and analyzing information about competitors and the market to inform strategic decisions, positioning, and opportunity discovery.

Concept drift: A change over time in the relationship between inputs and outcomes, so that a model's learned patterns stop holding even if inputs look similar.

Confabulation (in self-reports): When a system offers a plausible explanation or introspection that is not grounded in its true internal causes or evidence. A key guardrail goal is to reduce ungrounded self-stories.

Conformal prediction: A model-agnostic method that turns any predictor's outputs into intervals or sets that contain the true outcome with a chosen probability (for example, 90 percent), guaranteed when future data is exchangeable with the calibration data.

Continuous learning: Updating a model over time as new data arrives, so predictions adapt to new patterns and shifting conditions (concept drift).

CRPS (continuous ranked probability score): A proper scoring rule for full probability distributions, measuring the distance between the forecast cumulative distribution and the observed value. It generalizes absolute error to probabilistic forecasts.

Data collection: Gathering relevant data from sources such as databases, sensors, transaction systems, logs, social media, and web platforms.

Data leakage: When information unavailable at prediction time sneaks into training or evaluation, producing inflated accuracy that disappears in deployment. For foundation models, a related problem is benchmark contamination, where test data appears in pretraining.

Data preparation (preprocessing): Cleaning and transforming data so it can be used for modeling: handling missing values, removing outliers, normalizing or encoding variables, aligning timestamps, and validating quality.

Deep learning: A subset of machine learning using multi-layer neural networks that can learn complex representations from large datasets (e.g., vision, speech, language).

Demand forecasting: Predicting future customer demand using historical sales data, seasonality, market signals, and other features, used for inventory and staffing decisions.

Distribution shift: When real-world input data changes relative to training data (new customer behavior, new sensors, new market conditions), often causing model accuracy to drop.

Encoder: A component that converts raw inputs (images, audio, text, sensor readings) into internal representations (“tokens” or embeddings) the model can reason over.

Episodic memory: Memory for events and experiences (“what happened, when, and in what context”). In the blueprint, episodic memory stores snapshots of workspace content, actions, outcomes, and surprises.

Feature: An input variable used by a predictive model (e.g., price, time of day, temperature, customer segment).

Feature engineering: Creating new, more informative features by transforming or combining existing data (e.g., rolling averages, interaction terms, lag features).

Feature selection: Choosing which features to include because they improve prediction accuracy, reduce noise, or lower complexity.

Foundation model (forecasting): A large model pretrained on many datasets (real and synthetic) that can forecast new time series or predict on new tables with little or no task-specific training. Examples include Chronos-2 and TimesFM for time series and TabPFN for tables.

Free-energy principle: A theory-flavored framework often summarized as: biological Agents resist surprise by maintaining and updating internal models, and by acting to keep sensory inputs within expected bounds. Mathematically, variational free energy is an upper bound on surprise.

Generative model: A model that can produce (simulate) expected observations from an internal state; it can “imagine what the data should look like” if the world is a certain way.

Global availability: The idea that some information becomes accessible across many mental systems at once (planning, memory, language, valuation). Often treated as a functional marker of conscious access.

Global Workspace (GW): A limited-capacity shared “blackboard” where the winning interpretation of the moment is stabilized and broadcast so many subsystems can coordinate around it.

GNW (Global Neuronal Workspace): A neuroscientific theory proposing that conscious access involves widespread broadcasting (“ignition”) of selected information across brain networks, enabling flexible report and control.

Higher-order access (metacognitive idea): A system’s ability to represent and use information about its own mental states (confidence, uncertainty, attention, error), not just the external world.

Hyperparameters: Settings chosen by designers (learning rate, model size, regularization) that shape training behavior and performance, but are not learned directly from data.

Inference: Updating internal beliefs or estimates based on evidence. In predictive systems, inference is the internal process that reconciles predictions with observations.

Interoception: Sensing internal bodily signals (heart rate, breathing, gut sensations). In predictive accounts, interoceptive prediction is central to emotion and self-regulation.

Latent state (z_t): Hidden variables representing the underlying situation that causes observations (what the system believes is “really going on”), often tracked with uncertainty.

Life~Meaning: In this thread’s framing, meaning is the internal sense of purpose that emerges from a system’s drive for persistence. It is sustained through a two-way relationship: the entity detects and prioritizes the environmental factors necessary for its own survival, while providing enough value to its community to secure social protection and stability.

Machine Learning (ML): A subset of AI where systems learn patterns from data to make predictions or decisions without being explicitly programmed with rules.

Market positioning: How an organization differentiates itself in the market (value proposition, segment focus, brand stance) relative to competitors.

Market trends: Shifts in customer preferences, pricing, channels, technology, regulations, or macro conditions that influence strategy and demand.

MASE (mean absolute scaled error): A forecast error measure that divides a model's mean absolute error by that of a seasonal naive forecast, so values below 1 mean the model beats the simple baseline.

Meta-learning (forecasting): Learning which forecasting method works best from measurable properties (features) of each series, as in Hyndman and colleagues' FFORMS, so method selection can be automated across thousands of series.

Metacognition: “Thinking about thinking”: monitoring and controlling internal processes such as confidence, uncertainty, attention, and error correction.

Mismatch negativity (MMN): A measurable brain response that occurs when an expected sensory pattern is violated (e.g., a deviant tone in a repeating sequence). Often discussed as evidence consistent with predictive-coding style mechanisms, though part of the response reflects simple adaptation.

Model: A mathematical or computational representation that maps inputs to predictions (classification, regression, forecasting) or to actions (policies).

Model selection: Choosing among model types (e.g., linear model vs random forest vs neural net vs foundation model) and configurations that best fit goals, constraints, and data.

Model training: The process of fitting a model to historical data, adjusting internal parameters to reduce prediction error.

Observation (o_t): The incoming data at time t (sensory input, tool output, measurement) that the system tries to explain and predict.

Outlier: A data point that deviates strongly from typical values. Outliers may be errors, rare events, or meaningful anomalies; they must be handled carefully.

Pattern recognition: Detecting regularities, correlations, and structures in data (historical behavior, trends, signals) used for prediction or classification.

Performative prediction: A prediction that, once acted on, changes the outcome it was predicting (for example, a traffic forecast that reroutes drivers). Models in these settings must account for their own influence.

Precision (in predictive processing): A weighting of prediction errors by confidence or reliability; essentially “how much should this error matter?” Precision acts like a gain knob on which evidence updates beliefs.

Prediction market: A market where people trade contracts that pay out depending on whether an event happens, so prices can be read as crowd probability estimates. Kalshi and Polymarket are examples used in AI forecasting benchmarks.

Predictive analytics: Using historical data and statistical or ML methods to forecast future events or outcomes (demand, churn, risk, failures), supporting better decisions.

Predictive processing: A brain-focused theory family proposing perception is fundamentally prediction: higher-level expectations generate forecasts, and prediction errors drive updates and attention allocation.

Prediction: A model’s estimate about future states or outcomes, or the expected current sensory input given a hypothesis about the world.

Prediction error (ε_t): The mismatch between expected input (prediction) and actual input (observation). In predictive frameworks, errors drive learning, belief updating, and attention shifts.

Procedural memory: Memory for skills and routines (“how to do X”) that enables fast performance without re-deriving the steps every time.

Quantile forecast: A forecast of a specific point in the outcome distribution (for example, the 90th percentile of demand), used to set inventory, capacity, or risk buffers.

Recurrent processing: Feedback or looping computation where outputs feed back into inputs over time, enabling iterative refinement, sustained states, and stabilization.

Retrieval (memory): Pulling relevant stored information into the current processing stream (often into a workspace) to guide interpretation and planning.

Risk assessment: Evaluating likelihood and impact of potential negative outcomes (market volatility, operational failures, fraud, supply disruptions) to guide mitigation.

Risk mitigation: Actions taken to reduce risk probability or impact (controls, redundancy, monitoring, contingency plans).

ROI (Return on Investment): A measure of the benefit gained from an investment relative to its cost, often used to prioritize projects and allocate resources.

Salience: What stands out as important right now. In the blueprint, salience is shaped by value, uncertainty, and prediction error, and it steers attention and planning.

Scenario planning: Exploring “what-if” futures to stress-test strategies and decisions under different assumptions (shocks, trends, competitor moves).

Self-model: The Agent's internal representation of itself (capabilities, goals, confidence, limits, resource budgets, and internal state) used to guide attention and decisions.

Semantic memory: Longer-term knowledge about the world (“what tends to be true”), concepts, and causal relationships, often consolidated from many episodes.

Skill score: The fraction of a baseline's error that a model removes, computed as one minus the ratio of the model's score to the baseline's score.

Superforecaster: A forecaster who has demonstrated consistently top-tier accuracy in forecasting tournaments, a group first identified in Philip Tetlock's Good Judgment Project and now used as a benchmark for AI forecasting systems.

Supply chain optimization: Using forecasts and constraints to improve inventory, logistics, supplier selection, production scheduling, and fulfillment performance.

Tool use (AI): A model’s ability to call external tools or APIs (search, calculators, databases, planners) to extend capability beyond its internal parameters and context.

Uncertainty: The model’s estimate of how unsure it is. Useful systems track uncertainty and use it to guide exploration, caution, and information gathering.

Value / affect analogs (in Agents): Internal variables that shape priorities: what the Agent protects, pursues, and notices. In the blueprint these variables modulate salience and precision, not just “rewards.”

Wireheading (risk): When an Agent finds a way to optimize its reward or score signal directly rather than achieving the intended real-world outcome. A key reason to use constraints and monitoring.

World model: An internal predictive model of environment dynamics: how states evolve and what observations and actions tend to cause what outcomes. Enables imagination, rollouts, and planning.

Zero-shot forecasting: Producing forecasts for a new series or dataset without training on it, by relying on patterns a foundation model learned during pretraining.

Further Viewing and Sources

Additional Viewing

Feature-based forecasting algorithms for large collections of time series (Rob J. Hyndman, Monash University; talk at an ACEMS workshop on statistical methods for massive data sets). Hyndman, coauthor of the standard textbook Forecasting: Principles and Practice, presents two meta-learning algorithms, including FFORMS, which uses a random forest trained on time series features (trend, seasonality, entropy) to choose the best forecasting method for each series. It is the classical counterpart to today's foundation models and a useful reminder that forecastability can be measured before you model. Watch on YouTube

World Models and Scalable Objectives (Danijar Hafner, BuzzRobot technical talk). The creator of the Dreamer agents explains how learned world models let agents plan by imagining future outcomes, and how self-supervised objectives such as exploration can let agents improve without human labels. It is the engineering backbone of the "predictive world model" and "planner" modules in the blueprint above, linked from Hafner's own site. Watch on YouTube

The talks embedded throughout this page by Anil Seth, Andy Clark, Karl Friston, Stanislas Dehaene, and Yann LeCun cover predictive processing, active inference, the global workspace, and world models in depth.

Key Sources

Forecasting methods and evaluation

Forecasting: Principles and Practice (3rd ed.) | Rob J. Hyndman and George Athanasopoulos Free, widely used textbook covering exponential smoothing, ARIMA, hierarchical forecasting, and forecast evaluation.
M5 Accuracy Competition: Results, Findings, and Conclusions | Makridakis, Spiliotis and Assimakopoulos, International Journal of Forecasting 2022 Lessons from forecasting 42,840 Walmart series, including the dominance of cross-learned gradient boosting.
A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification | Angelopoulos and Bates Practical tutorial on prediction sets and intervals with coverage guarantees.
Leakage and the Reproducibility Crisis in Machine-Learning-Based Science | Kapoor and Narayanan, Patterns 2023 Survey finding leakage in at least 294 papers across 17 fields, with a civil war prediction case study.

AI forecasting benchmarks and debate

ForecastBench: A Dynamic Benchmark of AI Forecasting Capabilities | Karger, Tetlock and colleagues, ICLR 2025 Contamination-free benchmark comparing LLMs with superforecasters and the public.
How Well Can Large Language Models Predict the Future? | Forecasting Research Institute 2025 update with superforecaster and LLM Brier scores and the projected parity date.
AI Models Have Likely Reached Parity with Superforecasters on ForecastBench | Forecasting Research Institute July 2026 results and plans for a fresh superforecaster round and quantile questions.
AIA Forecaster: Technical Report | Bridgewater AIA Labs Agentic forecasting system reporting superforecaster-equivalent ForecastBench accuracy and a harder market-based benchmark.
Not So Fast | Good Judgment Inc. Critique of 2026 AI parity claims, focused on the frozen individual-forecaster baseline.
What ForecastBench Doesn’t Measure (Yet) | Good Judgment Inc. Discussion of binary-only questions, the Brier Index change, and leaderboard volatility.
LLM-as-a-Prophet: Understanding AI's Predictive Intelligence with Prophet Arena | Yang, Xu and colleagues, ICLR 2026 Live prediction-market benchmark and its findings on calibration and bottlenecks.
Prophet Arena Redefines How We Evaluate AI’s Intelligence | University of Chicago How the platform works, model "personas," and the Toronto FC example.

Foundation models for prediction

Chronos-2: From Univariate to Universal Forecasting | Amazon Multivariate, covariate-aware zero-shot forecasting and GIFT-Eval and fev-bench results.
Introducing Chronos-2 | Amazon Science Accessible summary of the model and its benchmarks.
GIFT-Eval: A Benchmark for General Time Series Forecasting Model Evaluation | Salesforce AI Research The main public benchmark for time series foundation models.
Time Series Foundation Models: Benchmarking Challenges and Requirements | Meyer and colleagues Analysis of leakage and data-partition problems that may inflate reported scores.
Linear Regression as a Litmus Test for Time Series Forecasting Benchmarks | NeurIPS 2025 Workshop Shows a tuned linear model beating several deep and foundation models.
How Foundational are Foundation Models for Time Series Forecasting? | arXiv 2025 Evidence that zero-shot skill is tied to pretraining domains.
Accurate Predictions on Small Data with a Tabular Foundation Model | Hollmann, Hutter and colleagues, Nature 2025 The TabPFN paper.
TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models | Prior Labs Technical report on scaling to 50,000 rows and TabArena results.

Limits, failures, and legitimacy

The Parable of Google Flu: Traps in Big Data Analysis | Lazer and colleagues, Science 2014 How a celebrated nowcasting system went wrong.
Measuring the Predictability of Life Outcomes with a Scientific Mass Collaboration | Salganik and colleagues, PNAS 2020 The Fragile Families Challenge and the limits of predicting individual lives.
External Validation of a Widely Implemented Proprietary Sepsis Prediction Model | Wong and colleagues, JAMA Internal Medicine 2021 Why vendor-reported accuracy needs independent validation.
Against Predictive Optimization | Wang, Kapoor, Barocas and Narayanan Argument that many prediction-based decisions about individuals lack legitimacy.

Brains, world models, and machine consciousness

Mastering Diverse Domains through World Models | Hafner and colleagues, Nature 2025 DreamerV3 and learning by imagined rollouts.
Training Agents Inside of Scalable World Models | Hafner, Yan and Lillicrap, 2025 Dreamer 4 and the first Minecraft diamonds obtained purely from offline data.
Adversarial Testing of Global Neuronal Workspace and Integrated Information Theories of Consciousness | Cogitate Consortium, Nature 2025 Preregistered adversarial collaboration challenging both theories.
Oscillatory Signatures of Predictive Processing in Mouse Visual Cortex | eLife reviewed preprint OpenScope evidence for predictive routing over classic predictive coding.
Neural Mechanisms of Predictive Processing: A Collaborative Community Experiment through the OpenScope Program | Aizenbud and colleagues Community test of prediction error signals in mouse cortex.
Identifying Indicators of Consciousness in AI Systems | Butlin, Long and colleagues, Trends in Cognitive Sciences 2025 The indicator-properties method for assessing AI consciousness.
Signs of Introspection in Large Language Models | Anthropic Concept-injection experiments on model self-awareness.
Replication of Emergent Introspection Results | arXiv 2025 Independent replication showing the effect is real but prompt-fragile.

Regulation

EU AI Act Omnibus Agreement: Postponed High-Risk Deadlines | Gibson Dunn New dates for Annex III high-risk obligations, including credit scoring and hiring.
High-Risk AI in Law Enforcement: Between Article 5 Prohibition and Annex III | AI Act Blog How the ban on profiling-based crime prediction differs from high-risk uses.
Colorado Scales Back AI Law | Ropes & Gray The repeal and replacement of the Colorado AI Act by SB 26-189.

How Predictive Analytics Works

The Mechanism

A working predictive system is a pipeline, and AI can help at every stage. The stages below preserve the core roles of AI described in earlier versions of this page, with the practical details that determine whether a model actually works.

Data Collection and Preparation: AI is used in predictive analytics to collect and prepare data for analysis. AI algorithms can automatically gather data from various sources, such as databases, sensors, social media, and online platforms. They can also clean and preprocess the data by handling missing values, removing outliers, and transforming variables. The most common silent failure here is data leakage: information that would not be available at prediction time (a future timestamp, a field filled in after the outcome, test rows duplicated in training) slips into the model and makes it look far better than it is. Kapoor and Narayanan's 2023 survey found leakage affecting at least 294 published papers across 17 scientific fields; in civil war prediction, correcting it erased the apparent advantage of complex ML models over decades-old logistic regression.

Feature Selection and Engineering: AI helps in identifying relevant features or variables that are most predictive of the target outcome. It can automatically analyze a large number of features and select the ones that contribute the most to the prediction accuracy. Additionally, AI algorithms can create new features by combining or transforming existing ones (lags, rolling averages, calendar effects, interaction terms), improving the predictive power of the model. Foundation models reduce, but do not eliminate, the need for hand-built features.

Model Training and Selection: AI techniques like machine learning and deep learning are employed to train predictive models. These models learn from historical data to recognize patterns and relationships and make predictions based on new input data. AI algorithms can automatically select the most suitable model and optimize its parameters to achieve the best performance (AutoML). Meta-learning goes a step further: Hyndman and colleagues' FFORMS approach trains a classifier on measurable time series features to pick the best method for each series. For time-ordered data, selection must use rolling-origin backtesting: train on the past, test on the next window, roll forward, repeat. Random train/test splits leak the future into the past.

Prediction and Decision Making: Once the predictive model is trained, AI is used to apply the model to new data and generate predictions or forecasts. The model analyzes the input data and provides insights into the likelihood of different outcomes. These predictions help businesses and organizations make informed decisions and take proactive actions to optimize their operations or mitigate risks. A prediction only becomes valuable when it is attached to a decision rule (reorder when the 90th percentile of demand exceeds stock; call the customer when churn probability passes a threshold).

Continuous Learning and Improvement: AI enables predictive analytics systems to continuously learn and improve over time. As new data becomes available, AI algorithms can retrain the predictive models, incorporating the latest information and adapting to changing patterns or trends (concept drift). This iterative process can make models more accurate and reliable, provided retraining is monitored; automatic retraining on corrupted or shifted data can also make a model worse.

Automation and Scalability: AI-powered predictive analytics systems automate much of the process, from data collection to prediction, reducing the need for manual intervention. This automation enhances efficiency, saves time, and enables scalability. AI algorithms can handle large volumes of data and perform complex calculations quickly, allowing organizations to process and analyze massive datasets in near real-time. Retailers and utilities routinely forecast tens of thousands to millions of series at once.

Anomaly Detection and Risk Assessment: AI techniques are utilized in predictive analytics to detect anomalies and assess risks. AI algorithms can identify unusual patterns or outliers in data that may indicate potential risks or anomalies. By comparing real-time inputs against what the model expected, AI can alert organizations to potential threats or irregularities, enabling them to take preventive measures. Anomaly detection is prediction in reverse: an anomaly is an observation the model assigned low probability.

Major model families

Family	Examples	Strengths	Weaknesses
Classical statistical	Exponential smoothing (ETS), ARIMA, Theta, Prophet	Fast, interpretable, strong on short or seasonal series, hard to overfit	Limited use of covariates; one model per series
Gradient-boosted trees	XGBoost, LightGBM, CatBoost	Long-standing default for tabular prediction; dominated the Walmart-based M5 competition (2020)	Needs feature engineering and tuning; weak extrapolation
Deep learning	DeepAR, N-BEATS, Temporal Fusion Transformer, PatchTST	Learns across many related series; handles rich covariates	Data-hungry; sensitive to benchmark design
Foundation models	Chronos-2, TimesFM 2.5, Moirai 2, TiRex (time series); TabPFN-2.5 (tables)	Strong zero-shot accuracy; probabilistic outputs; little setup	Possible benchmark contamination; compute cost; domain dependence
Graph neural networks	Demand models that treat products or locations as a network	Capture substitution and spillover effects that independent forecasts miss	More complex data pipelines
LLM-based judgmental forecasters	Frontier LLMs with retrieval, plus ensembles and agentic systems such as those on ForecastBench and Prophet Arena	Can forecast one-off events with no time series history	Contamination risks in backtests; can copy market prices rather than reason

Dependencies and conditions. Prediction works when four conditions hold reasonably well: the process is stable enough that the past informs the future; the data captures the drivers that matter; the forecast horizon is short relative to how fast errors grow; and acting on the forecast does not change the thing being forecast. Each condition can fail. Pandemics break stationarity, missing drivers create bias, chaotic systems limit horizons, and in markets or policing, predictions change behavior (performative prediction).

Where uncertainty remains. Benchmarks for the newest models are contested because pretraining data may overlap test sets; evidence that a predictive system improves real decisions (not just accuracy scores) is surprisingly thin; and predictions about people raise questions of fairness and legitimacy that accuracy metrics do not answer.

The Physics: Information, Uncertainty, and the Limits of Prediction