Market Trading

From PRIMO.ai
Jump to navigation Jump to search

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


AI market trading is the use of machine learning style methods, and increasingly large language models (LLMs), to decide what to buy or sell, how much, and how to execute the order in stocks, bonds, commodities, currencies, and crypto-assets. It is a branch of Predictive Analytics and Forecasting, but with a twist that makes it uniquely hard: the thing you are predicting is a price set by millions of other participants, many of them also using computers, who are actively trying to profit from the same information you have.

The field has deep roots. Quantitative investing grew from Harry Markowitz's portfolio theory (1952), the efficient market hypothesis of Eugene Fama (1970), and the factor models that followed. Firms such as Renaissance Technologies, D. E. Shaw, AQR, and Two Sigma industrialized systematic trading from the 1980s onward, and electronic markets turned speed into a business: high-frequency trading (HFT) firms now handle roughly half of U.S. equity volume as market makers and arbitrageurs. Academic work by Shihao Gu, Bryan Kelly, and Dacheng Xiu (2020) showed that neural networks and tree models can predict the cross-section of stock returns better than classic regressions, while Marcos López de Prado documented why most machine learning funds still fail. Since 2023, the arrival of ChatGPT and other LLMs added something new: models that can read news, filings, and earnings calls and turn them into trading signals or even place trades through brokerage APIs such as Alpaca's.

The central mechanism is a pipeline: data (prices, order books, fundamentals, news, alternative data) feeds signals (a forecast of return or risk), which feed portfolio construction (how much to hold given risk and constraints), which feeds execution (how to trade without moving the price against yourself), all wrapped in risk controls and monitoring. AI can help at every stage, but the pipeline only makes money if the signal's edge survives transaction costs, competition, and the statistical trap of finding patterns in noise.

What has changed is not that machines trade (they have for decades) but who can build them and what they can read. A hobbyist can now wire an LLM to a commission-free brokerage API in an afternoon, as the tutorials preserved below show. The evidence says that is the easy part. The hard part, and the focus of this page, is knowing whether a strategy has a real edge before money is on the line. Prediction markets, covered in their own section below, offer a different way to convert information into prices: event contracts that can also be assessed as probabilistic forecasts.


In markets, a prediction is only worth what is left after costs, competition, and the chance you fooled yourself. AI changes how fast edges are found, and how fast they disappear.


How AI Trading Works

The Mechanism

1. Data. Trading systems ingest market data (trades, quotes, full limit order books), fundamental data (earnings, balance sheets), macro data, and "alternative" data (news, social media, satellite imagery, web traffic, credit card panels). For text, LLMs have largely replaced word-count sentiment dictionaries. The critical discipline is point-in-time data: every input must be exactly what was knowable at that moment, or the backtest will cheat.

2. Signals (alpha models). A signal is a forecast, usually of the next period's return relative to peers, sometimes of volatility or of the probability of an event. Common families:

  • Cross-sectional machine learning: rank thousands of stocks each month using dozens to hundreds of characteristics (momentum, value, size, liquidity, volatility). Gu, Kelly, and Xiu found trees and neural networks win mainly by capturing nonlinear interactions between predictors.
  • Time-series and market timing: forecast an index, currency, or commodity. Kelly, Malamud, and Zhou argue that heavily parameterized models do better here than intuition suggests (see Recent Developments for the debate).
  • Event and news-driven: read headlines, filings, or transcripts and react. This is where LLMs show the clearest measured skill.
  • Market making and HFT: quote on both sides of the order book and earn the bid-ask spread, or exploit stale quotes across venues. These strategies live on microseconds and are driven by latency and microstructure models, not LLMs.
  • LLM and multi-agent systems: frameworks such as TradingAgents assign LLMs roles (fundamental analyst, sentiment analyst, bull and bear researchers, risk manager, trader) that debate before acting.

3. Portfolio construction and sizing. Signals are combined and converted into positions subject to risk limits, sector and factor exposure limits, leverage, and liquidity. Position sizing matters as much as signal quality; the Kelly criterion (below) explains why overbetting a real edge can still lose money.

4. Execution. Large orders are sliced over time with algorithms such as VWAP (volume-weighted average price) or TWAP (time-weighted), routed across exchanges by smart order routers, and increasingly optimized with reinforcement learning. John Hull and colleagues have shown reinforcement learning can find hedging strategies that cut transaction costs relative to textbook delta hedging. Poor execution silently erases thin edges.

5. Risk controls and monitoring. Professional systems run pre-trade checks (maximum order size, price collars, position and loss limits), kill switches, and post-trade surveillance. In the United States, broker-dealers with market access must maintain such controls under the SEC's Market Access Rule (Rule 15c3-5, 2010); in the EU, MiFID II requires testing, kill functionality, and controls for algorithmic traders. Monitoring also watches for regime change: a model trained in calm markets can behave very differently in a crash.

6. Research process. Every serious shop runs a pipeline of idea generation, point-in-time data assembly, backtesting with realistic costs, out-of-sample and paper trading, small live allocation, then scaling. López de Prado's critique of most failed ML funds is that individuals working alone repeat this loop until something "works" in-sample, which almost guarantees overfitting.

Dependencies and conditions. An edge requires that (a) the signal contains information not already in prices, (b) the information is acted on before others act on it, (c) costs are smaller than the edge, and (d) the pattern persists long enough to be harvested. AI helps most with (a), sometimes with (b), and does nothing for (c) and (d) unless the system is engineered for them.

Where uncertainty remains. How much of the published ML outperformance survives real costs at scale; whether LLM signals keep working as more traders use the same models; and how AI agents behave when they interact with each other in live markets.

Edges, Noise, Costs, and Risk

Markets have their own "physics": quantitative relationships that explain why trading is hard and why most impressive backtests disappoint.

1. Efficiency and the scarcity of signal. If a predictable pattern were easy to exploit, traders would trade it away. Grossman and Stiglitz (1980) showed markets cannot be perfectly efficient, because then no one would pay to gather information, but they can be efficient enough that the remaining signal is tiny. In Gu, Kelly, and Xiu's study, the best models explained well under 1 percent of the variance of monthly individual stock returns out of sample. That sounds useless, yet applied across thousands of stocks it was economically large, in some cases doubling the performance of leading regression strategies.

2. The fundamental law of active management. Grinold's law says an active strategy's information ratio is roughly IR ≈ IC × √BR, where IC (information coefficient) is the correlation between forecasts and outcomes and BR (breadth) is the number of independent bets per year. A tiny IC of 0.02 applied to 500 independent bets gives IR ≈ 0.02 × 22.4 ≈ 0.45, a respectable result. This is why quants trade many small bets, and why a single LLM making a few concentrated calls has little breadth to work with.

3. Sharpe ratios take years to verify. The Sharpe ratio is SR = (mean excess return) / (standard deviation). Its estimate from T years of data has a standard error of roughly √((1 + SR²/2) / T). To show a true annual Sharpe of 0.5 is different from zero at about two standard errors, you need around (2 / 0.5)² = 16 years of returns. A two-week trading competition, or a three-month backtest, says almost nothing about skill.

4. Multiple testing and backtest overfitting. Try enough strategies and one will look brilliant by chance. The expected best t-statistic among N skill-less strategies grows roughly like √(2 ln N); for 1,000 independent tries it is about 3.2, which on eight years of data looks like an annualized Sharpe ratio above 1. Harvey, Liu, and Zhu (2016) argued that new return factors should clear a t-statistic of about 3, not 2, for this reason. Bailey and López de Prado's deflated Sharpe ratio adjusts a backtest's Sharpe for the number of trials, the sample length, and fat tails. Publication itself is a test: McLean and Pontiff (2016) found anomaly returns were about 26 percent lower out of sample and about 58 percent lower after publication, consistent with both overfitting and traders arbitraging the effect away. More recent work by Jensen, Kelly, and Pedersen (2023) finds that most well-known factors do replicate once studied carefully, so the lesson is discipline, not nihilism.

5. Transaction costs and market impact. Every trade pays a spread, fees, and market impact. Empirically, impact follows a "square-root law": ΔP / P ≈ Y × σ × √(Q / V), where σ is daily volatility, Q is order size, V is daily volume, and Y is a constant near 1. Buying 1 percent of a day's volume in a stock with 2 percent daily volatility costs roughly 0.02 × √0.01 = 0.2 percent, or 20 basis points, before spread and fees. Net return equals gross edge minus turnover times cost per trade; many high-turnover ML signals look great gross and lose money net. Capacity follows directly: the bigger the fund, the larger Q, the more of the edge impact consumes.

6. Fat tails and volatility clustering. Daily returns are not normally distributed. Large moves follow a power law, P(|r| > x) ∝ x−α, with α near 3 for stocks (the "cubic law"), and volatility clusters: calm periods follow calm, turbulent follow turbulent (the pattern GARCH models capture). Models that assume normality badly underestimate crash risk, and an ML model trained mostly on calm data has seen few of the tail events that matter most.

7. Position sizing and the Kelly criterion. For a strategy with expected excess return μ and volatility σ, the long-run growth rate at leverage f is approximately g(f) = fμ − f²σ²/2, maximized at the Kelly fraction f* = μ / σ². Betting twice the Kelly fraction drives expected growth to zero, and more than that drives it negative, even with a genuine edge. Because μ is always estimated with error, practitioners bet "half Kelly" or less.

8. Speed and latency. At the HFT end, physics is literal. Light in optical fiber travels about a third slower than through air, which is why firms built microwave and millimeter-wave links between exchange data centers. Aquilina, Budish, and O'Neill (Quarterly Journal of Economics, 2022) found that latency-arbitrage races on the London Stock Exchange accounted for about a fifth of trading volume and acted like a small tax on liquidity. LLMs, which take hundreds of milliseconds to seconds per response, cannot compete on this axis; any edge they have lives at slower horizons.

9. Non-stationarity and reflexivity. Market relationships change as regulations, technology, and participants change, and trading on a pattern alters the pattern. Lopez-Lira and Tang report that the profitability of LLM news strategies fell as LLM adoption rose, the textbook signature of an edge being competed away (see feedback loops in stock market predictions). This is why models need monitoring and why yesterday's backtest is not tomorrow's forecast.

Experimental constraints. Finance has one history. You cannot rerun 2008 or 2020, so out-of-sample data accumulates at the speed of the calendar. Live, forward-looking tests (paper trading, small real allocations, live competitions) are the only clean experiments, and they are slow and noisy.

Why This Matters

Economic significance. Systematic and algorithmic strategies set prices and supply liquidity across global markets. Better forecasting and execution can lower trading costs for pension funds and improve price discovery. ML has become standard in risk modeling, execution, surveillance, and fraud detection, even where it does not pick stocks.

For individual traders, the evidence is sobering. Chague, De-Losso, and Giovannetti followed every person who started day trading Brazilian equity futures from 2013 to 2015 and persisted for at least 300 days: 97 percent lost money, and only a fraction of a percent earned more than a bank teller's wage, with no evidence of learning over time. Adding an LLM does not change the arithmetic of costs, competition, and noise; it mainly changes how quickly a trader can act on ideas, good or bad. The oldest AI-managed ETF, AIEQ (launched October 2017), is a public example: it beat the S&P 500 in some years but trailed the index over long stretches, for instance returning about 57 percent from launch to early 2022 versus about 84 percent for the S&P 500 total return index.

Fraud risk. "AI trading bot" is now a common pitch in investment scams. The CFTC's January 2024 advisory describes Mirror Trading International, a bot scheme that promised at least 10 percent per month and lost customers nearly 30,000 bitcoins, and reminds the public that AI cannot predict the future. Guaranteed or very high returns, pressure to recruit others, and unverifiable track records are red flags.

Market stability. Automation has produced spectacular failures: the May 6, 2010 Flash Crash, when the Dow fell nearly 1,000 points within minutes before recovering, and Knight Capital's August 1, 2012 software error, which lost about $440 million in roughly 45 minutes. Regulators now worry about AI-specific versions of these risks: many firms using similar models or the same foundation model could buy and sell in unison ("herding" or "model monoculture"), amplifying moves under stress. The IMF's October 2024 Global Financial Stability Report and the Bank of England's April 2025 Financial Stability in Focus both flag this, alongside concentration on a few AI providers and new cyber risks.

Fairness and market integrity. Simulation research by Dou, Goldstein, and Ji shows that reinforcement-learning trading agents can learn to sustain collusive, supra-competitive profits without communicating or intending to collude, harming price efficiency. Existing antitrust and manipulation law generally requires agreement or intent, so autonomous AI coordination is a genuine legal gray zone.

What would change if the optimists are right. Cheap, capable AI analysts would push prices to reflect public information faster, shrinking the easy edges in news and filings while lowering costs for everyone. What remains uncertain is whether that efficiency gain outweighs new fragility from correlated AI behavior, and whether today's measured LLM skill survives the competition it invites.

Recent Developments

Can LLMs predict returns? Evidence from news

Alejandro Lopez-Lira and Yuehua Tang (University of Florida) produced the most cited test. They asked GPT models whether each headline was good, bad, or neutral for a company's stock, using only headlines published after each model's knowledge cutoff to avoid contamination. Their latest version (October 2025) reports that GPT-4 captured the direction of the initial market reaction on about 90 percent of portfolio-days, but that reaction is too fast to trade on. The tradable part is the subsequent drift, which GPT-4 scores also predicted, especially for small stocks and negative news. Forecasting ability rose with model size; earlier models such as GPT-1, GPT-2, and BERT could not forecast returns. Crucially, strategy returns declined as LLM adoption rose, and for some topics (earnings reports, partnerships, clinical trial results) markets already reacted efficiently.

What this establishes: advanced LLMs extract real, economically meaningful information from news. What it does not: a durable edge for a retail bot. The signal is fast, concentrated in small and costly-to-trade stocks, and shrinking as others use the same tools.

LLMs trading live money: Alpha Arena

In October 2025, the startup nof1 gave six frontier models $10,000 each to trade crypto perpetual futures autonomously, with every trade public. Season 1 ran October 18 to November 3, 2025. Alibaba's Qwen 3 Max finished first at about $12,231 and DeepSeek second at about $10,489; Claude Sonnet 4.5 ($5,799), Gemini 2.5 Pro ($5,445), Grok 4 ($4,208), and GPT-5 ($4,126) all lost heavily. DeepSeek had been well ahead mid-season before a market downturn erased most of its gains. Season 1.5 switched to U.S. stocks, added several competition modes, and ended December 5, 2025, with an initially anonymous model later revealed as Grok 4.20 winning at about 12 percent on average over two weeks, ahead of GPT-5.1 and Gemini 3; within days, most models were again showing losses.

What this establishes: off-the-shelf LLMs given trading autonomy behave very differently (Qwen sized positions aggressively, others more cautiously), and most lost money in volatile conditions. What it does not: which model is a better trader. A few weeks of returns from one market path carries almost no statistical weight (see the Sharpe ratio arithmetic above), and the organizers themselves describe the contests as experiments under deliberately difficult conditions.

Multi-agent trading frameworks and live benchmarks

Research has moved from single prompts to agent teams. TradingAgents (Xiao and colleagues; oral presentation at an AAAI 2025 workshop) mimics a trading firm with analyst, researcher (bull versus bear debate), trader, and risk-management agents, and reports better cumulative return, Sharpe ratio, and drawdown than baseline strategies in backtests. Its open-source code became one of the most starred finance repositories on GitHub. Similar frameworks (FinMem, HedgeAgents, QuantAgents) report comparable backtest gains. The weakness of nearly all of them is evaluation: short historical windows, a handful of tickers, and LLMs that may have read about the very period being tested. That is why "live" benchmarks such as Agent Market Arena (October 2025), which evaluate agents in real time on verified data, matter more than backtests.

Machine learning return prediction: the complexity debate

Kelly, Malamud, and Zhou's "The Virtue of Complexity in Return Prediction" (Journal of Finance, 2024) proved that, contrary to the old rule of parsimony, models with more parameters than observations can deliver higher out-of-sample Sharpe ratios for market timing, and documented this in U.S. equity data. Subsequent papers challenged aspects of the result, arguing in effect that the gains depend on implicit regularization and small-sample behavior. In "Understanding the Virtue of Complexity" (2025), Kelly and Malamud responded that the challenges leave the core findings intact while clarifying the difference between "nominal" and "effective" complexity. The debate is ongoing; practitioners should read it as "large models can work with heavy regularization," not "bigger is always better."

AI collusion and herding

Dou, Goldstein, and Ji (NBER working paper, July 2025; revise-and-resubmit at the American Economic Review) placed reinforcement-learning speculators in a model market and found they autonomously sustained collusive profits through two mechanisms: price-trigger punishment strategies (which they call collusion through "artificial intelligence") and a generic learning bias they call "artificial stupidity." Prices became less informative. These are simulations, not evidence that collusion is happening today, but they define a concrete risk for regulators. Related experiments presented at central-bank conferences in 2025 and 2026 test whether LLM traders herd when shown other traders' actions, building on the IMF's finding that market participants rank herding and concentration as top risks of generative AI adoption.

Infrastructure: brokerages built for agents

Alpaca, the API-first brokerage featured below, launched an official Model Context Protocol (MCP) server in April 2025 (rebuilt as version 2 later) that lets assistants such as Claude, ChatGPT, and Gemini, plus coding tools, query markets and place stock, ETF, options, and crypto orders in natural language, with paper trading supported. Alpaca reports that API trading grew nearly fourfold in the first quarter of 2026, driven largely by AI agents, and it raised a $150 million Series D in January 2026 (at a reported $1.15 billion valuation) and a further $135 million in July 2026. This makes autonomous trading far more accessible, and it makes risk controls, toolset restrictions, and human approval steps more important.

Regulation and oversight

  • SEC. On June 12, 2025, the SEC withdrew its 2023 proposal on conflicts of interest in "predictive data analytics" used by broker-dealers and advisers, along with 13 other proposals. Existing fiduciary, best-interest, and anti-fraud rules still apply, and the SEC continues to bring "AI washing" cases against firms that misrepresent their use of AI (for example, the March 2024 settlements with Delphia and Global Predictions).
  • FINRA. Its 2026 Annual Regulatory Oversight Report (December 2025) added a dedicated generative AI section: firms should assess compliance obligations before deploying GenAI, keep human oversight, test for market abuse such as layering, spoofing, and wash trading, and treat autonomous AI agents as raising novel supervisory issues (tracking actions, restricting access).
  • CFTC. In January 2024 staff requested comment on AI in derivatives markets and issued the trading-bot scam advisory; a December 2024 staff advisory reminded registered entities that existing rules apply to AI use.
  • Europe and the UK. ESMA's May 2024 statement reminds investment firms that MiFID II obligations apply fully when using AI; the Bank of England's Financial Policy Committee is building ongoing monitoring of AI-related systemic risk.

Open questions

  • Will LLM news signals decay to zero as adoption grows, or will better models keep finding new information?
  • Can live, long-duration, statistically powered benchmarks replace short contests and contaminated backtests?
  • How should law treat collusion or manipulation that emerges from learning rather than intent?
  • Does widespread use of a few foundation models create correlated crashes, and can diversity of models be encouraged?


Stock Market Prediction with AI: Six Leading Platforms

AI assistants can help forecast a stock's direction, expected return, volatility, earnings response, or valuation range, but those are different prediction problems. An assistant that accurately explains why a stock rose yesterday has not demonstrated that it can predict tomorrow's return. The practical value of these platforms is in combining current research with structured data, reproducible calculations, independent challenges to an investment thesis, and a record of forecasts made before outcomes were known.

As of October 2026, six useful ecosystems to compare are OpenAI, Anthropic, Google, Microsoft, Perplexity, and xAI. The final two are selected for distinct research capabilities, not because a controlled benchmark establishes that they are the fifth- and sixth-best predictors: Perplexity specializes in source-backed web and financial-data retrieval; Grok offers direct analysis of X conversations and breaking social narratives. None has a demonstrated, general-purpose ability to reliably predict stock prices or earn excess returns after transaction costs.

Choose the Prediction Target Before Choosing a Tool

An actionable forecast specifies the security, information cutoff, target, time horizon, and baseline. Examples include:

  • Direction: probability that an adjusted closing price five trading days after the cutoff exceeds its adjusted closing price at the cutoff.
  • Excess return: predicted five-day stock total return minus the corresponding benchmark's total return.
  • Price range: 10th, 50th, and 90th percentiles of the stock's possible closing price at a stated future date.
  • Earnings reaction: probability of a greater than 5 percent close-to-close move on the first regular trading session after an earnings announcement.
  • Risk: estimated 20-trading-day volatility, probability of a specified drawdown, or expected shortfall under a stress scenario.

For a fair test, record the precise cutoff time and timezone, adjusted versus unadjusted prices, dividend treatment, trading calendar, price feed, benchmark, and handling of splits and delistings. Decide whether orders would hypothetically be filled at the next open or another realistic, available price. Otherwise, an apparent prediction can be a retrospective explanation or an impossible trade.

Provider Tools to use for stock prediction research Most useful contribution Critical caution
OpenAI ChatGPT research and data analysis; ChatGPT for Excel/Sheets; Codex and Responses API; financial-data plugins Structured company research, scenario modeling, repeatable quantitative analysis and backtests No market-data feed is guaranteed without an authorized integration; generated backtests need independent validation
Anthropic Claude; Claude for Excel; Claude Cowork and Claude Code; financial-data connectors Filings analysis, financial-model auditing, investment theses, automated research workflows A polished valuation is not a price forecast; connector coverage and entitlements vary
Google Google Finance with AI and Deep Search; Gemini; Google Sheets GOOGLEFINANCE; Gemini Enterprise for Financial Services Stock and macro research, earnings context, charting, accessible spreadsheet inputs GOOGLEFINANCE quotes can be delayed up to 20 minutes, and supported instruments vary
Microsoft Microsoft 365 Copilot in Excel; Python in Excel; Analyst; Researcher and Model Council; financial-data connectors Excel-based statistical analysis, forecasting, sourced research and independent review of assumptions Enterprise features and premium feeds may require licenses; Microsoft can use models from other vendors
Perplexity Perplexity Finance, Stock Screener, Research/Labs, Agent API finance search Fast source-linked comparisons of companies, earnings, consensus expectations and current headlines Always open the primary filing or transcript; source retrieval is not independently validated prediction skill
xAI Grok; Grok Web Search and X Search; Responses API Detecting and testing emerging news narratives and changes in social-market sentiment X discussions are nonrepresentative, manipulable and often reactive to price changes rather than predictive

OpenAI: ChatGPT, Research, Data Analysis, and Trading Agents

Tools: ChatGPT, Deep research, ChatGPT for Excel and Google Sheets, Public Equity Investing plugin, ChatGPT for Financial Services, and Responses API tools.

Conversational AI is useful in trading, but mostly not in the way the "ask ChatGPT what to buy" videos suggest. Where the evidence is strongest:

  • Reading and summarizing: filings, earnings call transcripts, central bank statements, and news, faster than any human team. Research shows advanced LLMs extract real signal from headlines (see Recent Developments).
  • Research assistance and coding: writing data pipelines, backtests, and broker API code, provided a human checks the logic. LLM-written backtests frequently contain look-ahead bias (using future data) and survivorship bias (testing only stocks that still exist).
  • Structuring a process: checklists, trade journals, and risk rules that reduce impulsive decisions.

Where it is weak or risky:

  • Predicting prices from charts or raw price series. There is little evidence LLMs beat simple baselines here, and they are prone to confident, plausible-sounding pattern stories.
  • Remembering the future. Asking a model to "backtest" a strategy over a period included in its training data invites contamination: it may already know what happened.
  • Autonomous execution without guardrails. Live contests show large losses from overconfident position sizing. Use paper trading, hard position and loss limits, and human approval for real orders.

Prompting tips for trading research (see Prompt Engineering (PE)): give the model point-in-time data rather than asking it to recall facts; ask for a probability and the reasoning, not a buy or sell instruction; ask it to list what would make the view wrong; and log every prompt and response so you can evaluate the model's track record honestly later.


How to use them:

  1. Collect evidence: use web-backed research or an authorized financial-data integration to retrieve the latest 10-K/10-Q, SEC filings, investor presentations, earnings transcripts, estimate revisions, company guidance and dated news. Have the assistant distinguish reported figures from analyst estimates and its own inferences.
  2. Build an explicit thesis: ask for bull, base, and bear cases, each tied to revenue growth, margins, valuation multiples, rates and identifiable catalysts. Request disconfirming evidence, not just arguments supporting a favorite stock.
  3. Analyze data: upload a point-in-time CSV of prices, returns, corporate actions and benchmark data. Ask data analysis to calculate momentum, rolling volatility, factor exposures, forward returns and confidence intervals. Export the underlying calculations and inspect the Python or spreadsheet formulas.
  4. Make the forecast testable: output the stock symbol, cutoff, horizon, probability of a positive return, expected excess return, 10th/50th/90th-percentile prices, source links and assumptions in a structured table. Freeze that forecast before new prices arrive.
  5. Automate carefully: use Codex to build a version-controlled data pipeline or use the Responses API for scheduled research and logging. A brokerage connection such as Alpaca's can be restricted to paper trading and read-only market information until tests justify any further access.

Example prompt: Using the attached point-in-time data and newly retrieved, cited filings, estimate the probability that [TICKER] will outperform the S&P 500 over the next 20 trading days. Timestamp every input; distinguish confirmed facts, current consensus and assumptions; report a baseline estimate, bull/base/bear scenarios, uncertainty, costs, and conditions that would invalidate the forecast. Do not trade.

Specialized 2026 offering: ChatGPT for Financial Services combines institution-oriented research and financial-model workflows with built-in licensed data for eligible organizations. ChatGPT for Excel and Google Sheets is available for workbook-centered analysis. Neither product supplies a verified trading edge by itself; access to premium feeds and tools depends on plan and organizational permissions.

Anthropic: Claude, Claude Code, and Financial Research Agents

Tools: Claude, Claude for Financial Services, Claude for Excel and finance skills, and Claude Cowork/Claude Code financial agents.

How to use them:

  1. Read dense disclosures: provide the last several quarters of filings and transcripts, then ask Claude to extract guidance changes, segment growth, nonrecurring charges, accounting-policy changes, share dilution, debt maturities and important caveats. Require page references and exact periods.
  2. Develop a valuation framework: in Claude for Excel, build or audit a three-statement model, discounted cash flow model, comparables, and a sensitivity table for discount rate, terminal growth, revenue and margins. Separate market price from estimated intrinsic value.
  3. Challenge the investment narrative: use a bull analyst, a bear analyst and a risk reviewer as separate research roles. Require each to cite contradictory evidence and to identify uncertainties that cannot be resolved by the documents provided.
  4. Create repeatable research: use Claude Code or Cowork finance workflows to fetch permitted data, refresh assumptions, flag material changes, produce a structured investment note and save a dated prediction. Maintain a human approval gate for each external or financial action.

Example prompt: Read these four quarterly filings and earnings transcripts for [TICKER]. Reconstruct the change in management guidance, compare current consensus with historical forecasting errors, model three valuation scenarios, and state a probability distribution for the next earnings reaction. Show which inputs come from primary documents and which are assumptions; identify what would reverse each conclusion.

Distinct strength: Claude's document-oriented workflow and Excel integration are useful for fundamental analysts who need to audit a financial model. Anthropic's 2026 financial-services release also describes ready-to-run Cowork and Claude Code agent templates and connectors to market-data services such as FactSet, S&P Capital IQ, LSEG, Morningstar and Financial Modeling Prep. These are integrations, not blanket entitlements to licensed data.

Google: Gemini, AI-Powered Google Finance, and Sheets

Tools: Google Finance, Google Finance Deep Search and earnings features, Gemini, GOOGLEFINANCE in Google Sheets, and Gemini Enterprise for Financial Services.

How to use them:

  1. Start with the market view: open the stock in Google Finance to review financials, recent news, price history, technical charts and the earnings calendar. Where available, use Deep Search to investigate a question that spans many companies or macroeconomic regimes.
  2. Separate signal from narrative: ask Gemini to compare reported revenue, earnings, forward guidance, expectations and interest-rate sensitivity. Inspect the sources behind each generated explanation.
  3. Make a live research sheet: use the GOOGLEFINANCE function for supported historical or quote fields, calculate daily returns and moving averages, and combine them with your own dated earnings and news observations. For example, =GOOGLEFINANCE("NASDAQ:GOOG","price",TODAY()-90,TODAY(),"DAILY") returns supported historical prices.
  4. Run a reproducible experiment: divide data by date, fit a transparent baseline such as momentum or a simple regression, then compare an AI-assisted signal on the identical forward period. Do not use future-revised fundamentals or undated search summaries in past forecasts.
  5. In an enterprise environment: use Gemini Enterprise for Financial Services and approved financial-data connectors to build governed, auditable research workflows where available.

Example prompt: Compare [TICKER] with three publicly traded peers. Using current, dated sources, isolate the roles of growth, valuation, interest rates and earnings revisions in recent performance. Design a simple five-day excess-return forecasting test that I can reproduce in Google Sheets, and list limitations of each data source.

Data caveat: Google's documentation says GOOGLEFINANCE quotes may be delayed by as much as 20 minutes, does not support all securities, and is not intended as a trading feed. Historical GOOGLEFINANCE results have restrictions when accessed through the Sheets API or Apps Script. The Google Finance AI interface and Gemini Enterprise offerings are distinct products with different feature availability. Google Finance also incorporates some prediction-market information, which should not be confused with a stock-price forecast.

Microsoft: Copilot in Excel, Analyst, Researcher, and Model Council

Tools: Copilot in Excel for finance, Python-powered Copilot analysis, Researcher, Analyst, and Researcher Model Council.

How to use them:

  1. Build the analytical workbook: import a dated price and fundamentals table, then ask Copilot in Excel to create an assumptions sheet, a return-and-risk dashboard, scenario tables, and an auditable forecast tab.
  2. Use Python where suitable: calculate correlations, distributions, rolling risk, walk-forward predictions and a simple backtest using the Python analysis capabilities available in Copilot in Excel. Check generated formulas and code rather than relying on a narrative description.
  3. Investigate drivers: ask Researcher to assemble a sourced report on earnings guidance, sector developments, regulation and competitor performance. Ask Analyst to investigate patterns in the numerical data.
  4. Compare independent explanations: if available, use Model Council within Researcher to contrast GPT- and Claude-based research reports, investigate their disagreements and challenge weak evidence. This is model disagreement analysis, not six fully independent data sources.
  5. Connect licensed feeds: enterprise finance connectors can bring services such as LSEG, Moody's, FactSet, Daloopa, Morningstar and S&P Global into Excel, subject to licenses, permissions and product availability.

Example prompt: In this workbook, estimate and document a 20-trading-day expected return and 90 percent prediction interval for [TICKER], using only observations available before [CUTOFF]. Compare a historical-mean baseline, momentum and one explicitly described regression. Create a walk-forward evaluation with transaction costs, report the error of each method, and highlight possible data leakage.

Distinct strength: Microsoft's advantage is a familiar Excel-centered decision surface where assumptions, formulas and results can remain visible. Finance-focused Copilot features announced in 2026 include editable finance models, configurable skills and financial-data connectors. Copilot's Researcher may use OpenAI and Anthropic models, so a comparison against ChatGPT or Claude isn't necessarily a comparison of unrelated underlying models.

Perplexity: Finance, Stock Screener, Research, and Finance Search

Tools: Perplexity Finance, Stock Screener, Earnings Calendar, Portfolio tools, and Perplexity API Platform.

How to use them:

  1. Screen candidates: start with the stock screener to narrow a universe by valuation, revenue growth, profitability, market capitalization or other supported fields. Record the filters and filing periods so later comparisons use the same definitions.
  2. Examine what is priced in: compare earnings expectations, reported results, company guidance, valuation, recent headlines and competing explanations. Follow citations to filings, earnings audio/transcripts and primary releases.
  3. Build an evidence ledger: for each potential catalyst, save the original source, publication time, consensus expectation, direction of surprise and confidence that the information was available before the intended prediction cutoff.
  4. Run a recurring comparison: use Research/Labs, or the Agent API's finance-search capability where available, to update a table of companies and identify genuine changes since the prior snapshot. Keep raw retrieved values and avoid silently replacing earlier estimates with revised figures.

Example prompt: For [TICKER] and its three closest comparables, retrieve the most recent earnings releases and filings. Compare actual results with dated consensus expectations, valuation and guidance revisions. Provide a source-linked table, a plausible five-day return range under three scenarios, and the strongest evidence against the central scenario.

Distinct strength: Perplexity Finance is a convenient first research stop because it combines a finance interface with cited external information, screeners, earnings materials and financial-data retrieval. The official platform also offers agent tools, including finance search. Coverage and timeliness differ by source; a citation only proves where a claim was found, not that a model has forecasted the market correctly.

xAI: Grok, X Search, and Breaking-News Sentiment

Tools: Grok, Grok Web Search, X Search, and Combining search tools.

How to use them:

  1. Watch a narrowly defined catalyst: use Grok to find recent discussion about an earnings release, regulatory action, product recall, litigation or supply-chain change. Search posts within an explicit time window and identify the original company or regulator source.
  2. Distinguish news from repetition: separate first-hand disclosures from reposts, bots, commentary and later reactions to the stock price. Ask for timestamps, quoted source links, original authors and duplicated narratives.
  3. Measure sentiment rather than guessing returns: classify a fixed sample of posts by event, topic, novelty and positive/negative interpretation. Compare the measure against subsequent returns only in a forward-tested study that controls for market and sector movement.
  4. Combine sources: with the API, use Web Search plus X Search to compare news reports with social discussion. Constrain dates and relevant accounts when possible, record search criteria and retrieved items, and calculate the effect of differing search windows.

Example prompt: Analyze public discussion of [TICKER] during [START] to [END], separating official disclosures, independently verified reporting, analyst commentary and unverified posts. Identify the first observed timestamp of each material claim; deduplicate repetitions; evaluate whether sentiment appears to lead or follow the price change. Do not turn unverified social volume into a return forecast.

Distinct strength: Grok and its API can examine the changing narrative on X in a way that is useful for studying attention, rumors and event diffusion. X Search supports date limits and account filtering. This is also its main limitation: social-media samples can be manipulated, selected by an opaque ranking process, and biased toward sensational activity. Search access, API billing and available features differ by product and account.

A Repeatable Workflow Across All Six

The most useful experiment gives each platform the same information at the same cutoff and asks the same measurable question. Without that discipline, the platform with the newest information can look like a better forecaster even when it is simply describing events the other models have not yet seen.

  1. Define a universe and horizon: for example, a fixed list of liquid U.S. stocks, forecasts recorded at 4:15 p.m. Eastern on each Friday, and returns through the next Friday's close. Document holidays, early closes and unusual corporate actions.
  2. Establish baselines: use a no-change or historical-frequency probability for direction, a benchmark return for excess returns, and a simple volatility forecast. AI needs to beat a defensible alternative, not just chance on one anecdotal winner.
  3. Prepare a common evidence packet: include the same date-stamped prices, filings, earnings information and macro indicators for all vendors. For a separate real-time-research test, allow their own tools but log the sources they retrieve; do not merge results from those two experimental designs.
  4. Demand machine-comparable outputs: symbol; forecast creation time; prediction cutoff; target date; probability of positive excess return; expected excess return; 10th/50th/90th-percentile price or return; main catalysts; source URLs; and reasons the model could be wrong.
  5. Freeze predictions before outcomes: save immutable CSV or JSON records, model/tool version, prompt and data-source timestamps. Do not retrospectively revise original forecasts.
  6. Score correctly: for binary direction, compute the Brier score (forecast probability minus outcome)2 and calibration by probability bucket. For returns, use mean absolute error, root mean squared error and directional accuracy. For probability distributions, test interval coverage. Evaluate results by market regime and against baselines.
  7. Test economic value separately: simulate only executable orders, with spread, slippage, brokerage fees, borrowing constraints, dividends, position limits, exposure controls and taxes where relevant. Report net return, volatility, drawdown, turnover, capacity and uncertainty. A forecast can be statistically useful but uneconomic to trade.
  8. Keep permissions bounded: separate read-only research credentials from trading credentials. Use paper accounts for automation experiments. Require explicit human authorization before transmitting real orders, and block trading on material nonpublic information.

Suggested forecast record:

Field Example or definition
Symbol and exchange NASDAQ:GOOG (illustrative; not a recommendation)
Prediction made at ISO 8601 timestamp, including timezone
Forecast horizon Five trading sessions after the stated cutoff
Target Probability stock total return exceeds benchmark total return
Probability and interval Numeric probability plus a specified prediction interval
Data provenance Dated source URLs, downloaded snapshots, and feed names
Model provenance Provider, specific model/tool, settings and prompt version
Outcome and scoring Realized target, Brier score or return error, and applicable costs

A useful practical division of labor: use Perplexity or Google Finance for initial sourced discovery; OpenAI or Claude for filings, analysis and hypothesis formation; Microsoft Copilot or Google Sheets for auditable quantitative work; and Grok for checking whether social narratives are influencing short-horizon market attention. Then reverse the assignments on selected cases to detect provider-specific failure modes. This is a workflow proposal, not a claim that the named vendors have superior forecasting performance in those roles.

Primary Sources and Technical Documentation


Prediction Markets

Prediction markets turn uncertain future events into tradable contracts. Rather than trading shares in a company, participants buy and sell claims tied to a specific outcome: a central bank interest-rate decision, an inflation reading, a product launch, a sports result, an election, or a geopolitical development. Their prices can act as continuously updated forecasts because participants have a financial incentive to correct prices they believe are wrong. They are also speculative financial products, and many people use them primarily for entertainment or gambling rather than forecasting.

The distinction from conventional market trading is the underlying payoff. A stock represents a claim on an enterprise; a binary event contract ordinarily settles to a fixed amount if a precisely defined event occurs, or to zero if it does not. That makes event prices easier to interpret as probabilities, but it does not make them infallible. The quality of the forecast depends on the contract's wording, market liquidity, competing information, trader incentives, and how the result will be verified.

Why this matters to AI: prediction markets create a measurable, continuously revised target for forecasting and predictive analytics. An AI assistant can monitor changing market odds, retrieve the evidence behind an event, produce an independent probability forecast, and compare its calibration with market prices. A trading agent can then decide whether any apparent discrepancy is large enough to survive fees, spread, slippage, and resolution risk. This is related to stock trading, but unlike predicting an open-ended price, a resolved binary market eventually provides an unambiguous answer against which a forecast can be scored.

How Prediction Markets Work

Event Contracts and Implied Probabilities

For an ordinary binary contract, a winning Yes share pays $1 at settlement and a losing share pays $0. A No share has the complementary payoff. For example, consider a hypothetical contract: Will a specified economic indicator exceed a stated threshold in its initial official release on a specified date?

  • If a Yes share trades for $0.62, its simple, fee-free implied probability is 62 percent. This is a market-based estimate, not a claim that the outcome has become certain.
  • If the event occurs, a share bought for $0.62 settles for $1, yielding $0.38 before costs. If it does not, the buyer loses the $0.62 paid.
  • If an independent, properly calibrated model estimates a 70 percent chance, the hypothetical expected profit from buying at $0.62 is 0.70 × $0.38 − 0.30 × $0.62 = $0.08 per share before fees and market impact.
  • If the price subsequently rises from $0.62 to $0.75, a position may be sold before resolution, provided there is a buyer and the exchange permits closing it. The sale profit is not guaranteed until an order executes.

A price is not always a probability. More precisely, a binary price approximates a risk-neutral, market-implied probability under assumptions about rational participants, no arbitrage, frictionless trading, and reliable settlement. A participant's desire to hedge a loss elsewhere, preference for long-shot bets, limited capital, transaction fees, and unequal access to information can all move the price away from a statistically calibrated probability. A quoted last trade can also be stale; executable bid and ask prices are often more useful.

Contract type Example What the market estimates Important complication
Binary (Yes/No) Will a published economic number exceed X? Probability of the defined event The definition of published, the cutoff, and revisions must be explicit
Categorical Which of several possible outcomes occurs? Distribution over mutually exclusive outcomes Outcome categories must cover all possibilities without overlap
Range or threshold Will inflation fall in a specified interval? Probability mass across ranges or thresholds Adjacent ranges must use consistent data sources and time windows
Conditional Will a measurable result occur if a specified policy is adopted? Conditional probability under the specified scenario Selection effects and the counterfactual problem complicate interpretation
Continuous-value design What value will an indicator take? A mean, median, quantile, or distribution, depending on contract design A single threshold probability is not the expected value of the indicator

Contract resolution is part of the product. Every contract needs an authoritative data source, deadline, timezone, treatment of delayed reports, rules for corrections and ties, and a documented dispute process. A question that sounds unambiguous in ordinary language can be extremely difficult to settle objectively. In decentralized markets, blockchain settlement does not eliminate the need for an external reporting mechanism, sometimes called an oracle, to establish what happened in the real world.

The Mechanism: From Beliefs to Prices

1. Market creation. An exchange or market organizer selects a forecastable question, specifies settlement rules, and identifies an independent source of truth. Questions need to be understandable, measurable, and legitimate to offer.

2. Liquidity. An order-book market matches buyers and sellers at bid and ask prices. Some platforms and experimental designs instead use automated market makers or scoring-rule-based liquidity subsidies. Thin order books may show attractive prices but prohibitively wide spreads or little capacity to trade.

3. Information aggregation. New evidence changes traders' private estimates. Those who consider the market too pessimistic buy Yes; those who consider it too optimistic sell Yes or buy No. Trading moves the marginal price. An effective market translates many dispersed observations into one interpretable signal, even when no single participant holds all the relevant information.

4. Updating and settlement. Prices change as announcements, injuries, official releases, or other information arrive. At the specified resolution, an adjudicator applies the contract rules and pays the winning side. Disputes, ambiguous wording, bad data, or a compromised oracle can delay or distort settlement.

5. Evaluation. The contract's final outcome is compared with prices recorded before resolution. Forecasts must be scored over many events, using the prices and information available at each forecast time. Selecting only successful trades or evaluating prices moments before settlement produces a misleading accuracy claim.

Automated market makers. A classic design from Robin Hanson is the logarithmic market scoring rule (LMSR). In a binary version, an automated market maker can use a cost function C(qYes,qNo) = b ln(exp(qYes/b) + exp(qNo/b)), where b controls liquidity and the cost of moving prices. The instantaneous Yes price is exp(qYes/b) / (exp(qYes/b) + exp(qNo/b)). Its advantage is that traders can transact even without a waiting counterparty, in exchange for a bounded-loss market-making subsidy. This is a foundational mechanism, not a claim that every modern commercial prediction exchange uses LMSR.

Forecast Accuracy and How to Measure It

The important question isn't whether a market called one high-profile event correctly. It is whether its probabilities are well calibrated across many events, time horizons, and conditions.

  • Calibration: Among events quoted at approximately 70 percent, roughly 70 percent should occur. A market that assigns 90 percent to events that happen only 60 percent of the time is overconfident.
  • Brier score: For a binary event, Brier = (p − y)2, where p is the probability from 0 to 1 and y is 1 for Yes or 0 for No. Lower is better. Average it across pre-specified forecasts and compare against appropriate baselines.
  • Log loss: Penalizes confident forecasts that are wrong more strongly than the Brier score. It is useful when very small or very large probabilities matter.
  • Resolution horizon: Record forecasts 30 days, seven days, and one day before settlement separately. A market can become highly accurate just before the event while still having little long-range predictive value.
  • Independence and representativeness: Contracts about a single contest may be strongly correlated. Treating thousands of related Yes/No derivatives as thousands of independent observations inflates apparent statistical certainty.
  • Economic value: A well-calibrated market can still be unprofitable to trade after fees. Conversely, a forecaster with useful probabilities may not have enough capital, liquidity, or speed to monetize the advantage.

In their foundational 2004 study, Justin Wolfers and Eric Zitzewitz found that prediction markets often aggregate dispersed information effectively and frequently outperform moderately sophisticated benchmarks. This is evidence for their usefulness, not a universal law about every market or topic. More recent evidence is more granular:

  • A February 2026 Federal Reserve research paper by Anthony Diercks, Jared Katz, and Jonathan Wright compared Kalshi macroeconomic contracts with conventional forecasts. The authors reported competitive forecasting performance and information-rich, continuously updated distributions, with useful results for federal funds rates and inflation. It is a research paper, not an official Federal Reserve endorsement of every contract.
  • A 2026 Kalshi-authored calibration study evaluated more than 2.2 million resolved markets and reported strong aggregate calibration close to settlement. Its scale is noteworthy, but it is platform-authored research, and performance far ahead of settlement and across individual categories still requires separate scrutiny.
  • A 2026 study in the International Journal of Financial Markets and Derivatives found that forecast performance varies by category, with evidence of distortions and negative average trader returns in some samples. This reinforces the distinction between price discovery and a profitable trading system.

Prediction Markets | Wolfers and Zitzewitz, NBER 2004 Foundational review of the theory and accuracy evidence.
Kalshi and the Rise of Macro Markets | Diercks, Katz and Wright, Federal Reserve 2026 Empirical comparison with macroeconomic forecasts.
Calibration in Prediction Markets: Theory and Evidence | Kalshi Research, August 2026 Platform-authored large-sample calibration analysis.
Prediction Market Efficiency: Evidence from Kalshi on Pricing Accuracy, Forecasting, and Risk | Tanay Subramanian, 2026 Peer-reviewed study of differences by category and forecasting horizon.

The Pentagon's 2003 Prediction Market Shutdown

FutureMAP and the Policy Analysis Market

One of the most consequential early prediction-market experiments never reached its intended public launch. In July 2003, the United States Department of Defense shut down the Policy Analysis Market (PAM), a proposed geopolitical forecasting exchange developed under the Defense Advanced Research Projects Agency (DARPA). PAM was part of Future Markets Applied to Prediction (FutureMAP), a research effort begun in 2001 to investigate whether market-based forecasts could improve strategic intelligence and reduce surprises.

The idea was to create contracts on measurable political, economic, and military developments in the Middle East. Areas under consideration included Egypt, Iran, Iraq, Israel, Jordan, Saudi Arabia, Syria, and Turkey. Rather than relying exclusively on intelligence analysts to produce narrative assessments, the project would test whether incentives and price movements could summarize information scattered across people and organizations.

An especially innovative element was the conditional event market. For example, a contract might ask how a specified measure of regional stability would change if a government took one policy action rather than another. Those two prices might expose assumptions about policy effects that were invisible in an ordinary forecast. They would still represent forecasts under scenarios, not proof of causation or an experimental estimate of what a policy would actually do.

Robin Hanson, an economist associated with the market's design and the development of market scoring rules, was among the researchers involved. Net Exchange developed PAM as a public-facing prototype. As Hanson's historical archive documents, a limited trial with about 100 participants was being planned for September 2003, with broader trading anticipated in January 2004. The planned public exchange was canceled before it opened. It is therefore incorrect to describe the episode as the closure of an already operating public Pentagon terrorism-betting exchange.

Why It Was Canceled in July 2003

On July 28, 2003, Senators Ron Wyden and Byron Dorgan publicly criticized PAM after its website and publicity materials attracted attention. Illustrative screens included references to an assassination of Palestinian leader Yasser Arafat, an attack involving North Korea, and the overthrow of Jordan's monarch. News coverage described a potential market in assassinations, terrorist strikes, and coups. The ensuing controversy raised several objections:

  • Moral legitimacy: Critics argued that wagering on political violence or atrocities was inappropriate for a government-funded project.
  • Perverse incentives: Allowing participants to profit from a violent event might create incentives to cause, accelerate, or influence it.
  • National security: Public trading on sensitive possibilities could communicate intelligence interests to adversaries or create opportunities to misuse confidential information.
  • Public accountability: Legislators questioned the value and political judgment behind spending defense research money on such an experiment.

On July 29, 2003, the Pentagon announced the project's termination in the face of the criticism. Deputy Defense Secretary Paul Wolfowitz said at a Senate hearing that the program was being terminated; DARPA Director Tony Tether also publicly confirmed it would not continue. The controversy contributed to the turmoil around DARPA's Information Awareness Office, directed by John Poindexter, who resigned shortly afterward.

Important historical distinction: PAM's designers disputed the characterization of their plan as a marketplace for individual terrorist attacks or assassinations. Hanson says the most inflammatory examples appeared in a faint background mock-up and were not intended as contracts for trading on specific attacks. A retrospective article published in the CIA's Studies in Intelligence likewise describes the program primarily as an experiment in forecasting strategic developments. Those accounts do not erase the legitimate ethical concerns prompted by the examples, but they explain why describing PAM simply as an operational market that invited people to bet on murders is misleading. Contemporary news reports document why the proposal produced such an immediate political response.

What was lost: The shutdown ended a proposed controlled test of whether incentive-based forecasts could complement intelligence estimates, including analyses of conditional policy scenarios. There is no valid record of PAM's intended public market performing better or worse than analysts, because the public experiment never ran. The strongest lesson is about institutional legitimacy and market design: even if a forecasting mechanism is mathematically interesting, its questions, incentives, participants, and potential social consequences must be acceptable to those affected.

The Policy Analysis Market and FutureMAP Archive | Robin Hanson, George Mason University Firsthand chronology, planned test, and designer's explanation of the disputed examples.
Using Prediction Markets to Enhance US Intelligence Capabilities | Puong Fei Yeh, Studies in Intelligence, 2006 Retrospective government-published research on FutureMAP's purpose and intelligence-market design.
Pentagon Drops Bid for Futures Market | The Washington Post, July 29, 2003 Contemporary account of the immediate cancellation and official statements.
Designing Real Terrorism Futures | Robin Hanson, Public Choice, 2006 Technical examination of the different, more controversial proposal to forecast attacks directly, which Hanson distinguishes from PAM's actual design.

A Striking 2026 Pentagon Parallel

On October 20, 2026, Defense Secretary Pete Hegseth issued a directive prohibiting service members and Defense Department personnel from using classified or other nonpublic government information to trade on outcomes involving military operations, including through platforms such as Polymarket and Kalshi. The memorandum became public on October 22, 2026. This was not a general ban on all prediction-market participation by everyone in the military, nor a new shutdown of a Pentagon-run exchange. It targeted exploitation of nonpublic official information.

The immediate context included an April 2026 prosecution in which a Special Forces soldier was accused of using advance knowledge about the operation to capture Venezuelan President Nicolás Maduro to profit through prediction-market contracts. The accusation was being litigated, and a criminal charge is not a conviction. Separate CFTC enforcement actions in 2026 also addressed alleged or established misuse of sensitive information on event-contract platforms. In August, for example, the CFTC announced a settlement involving a former White House teleprompter operator who had traded on words the president would use in speeches.

The concern in 2026 was somewhat different from that in 2003: a liquid public market can reveal a signal about operational plans when people who know sensitive information place trades. An unusual price move or cluster of wagers could be analyzed by an adversary. Financial stakes can also compromise professional judgment when a participant has some ability to influence the outcome being traded.

The two episodes are a useful comparison. In 2003, Congress and the public questioned whether the government should create markets related to dangerous geopolitical events at all. In 2026, the government faced questions about whether its own personnel were leaking or exploiting confidential information through privately operated markets. Both cases show that prediction markets are not merely forecasting dashboards. Their incentives and observable trading behavior can have consequences outside the market itself.

Hegseth Bans DoD Personnel from Using Nonpublic Info to Bet on U.S. Military Operations | CBS News, October 22, 2026 Report on the October 20 memorandum and its scope.
New Pentagon Directive Bans Troops from Using Nonpublic Info to Bet in Prediction Markets | Associated Press, October 22, 2026 Independent reporting on the directive and the earlier criminal case.
CFTC Orders Gabriel Perez to Pay $172,000 for Insider Trading of Mention Market Event Contracts | CFTC, August 28, 2026 Official enforcement case involving nonpublic speech information.

Platforms and Uses

Platform or model General approach Useful context
Kalshi U.S. federally regulated exchange listing event contracts, including macroeconomic and other events Check the specific contract rules, fees, and state-level legal issues
Polymarket Widely followed event market with blockchain-based infrastructure; U.S. and international products have distinct legal and operational arrangements In 2022, the CFTC penalized its earlier unregistered U.S. activity; the CFTC's later DCM registry lists QCX LLC doing business as Polymarket US
PredictIt Political forecasting platform associated with an academic research tradition Product offerings and U.S. authorization conditions have changed over time
Iowa Electronic Markets Academic real-money research markets dating to 1988 Historically important example of elections and information aggregation
Manifold Primarily play-money and community forecasting Useful for exploring question wording, smaller topics, and forecasting practice; not equivalent to a regulated real-money exchange
Metaculus Structured forecasting and aggregation of forecaster probabilities Not fundamentally an exchange of financial contracts; useful as a nonmarket comparison baseline

Practical applications: macroeconomic scenario tracking, election and policy forecasting, business launch deadlines, supply-chain and weather contingency planning, sports forecasting, intelligence analysis, research prioritization, and the design of risk hedges. Many useful forecasting questions do not attract enough speculative volume to support a viable commercial exchange. An organization may need to subsidize liquidity, reward accurate forecasts, or use noncash scoring instead.

Market data and APIs. For quantitative analysis, collect each contract's exact question, contract identifier, initial creation time, bid, ask, last trade, volume, open interest where reported, official resolution source, timestamped price history, and final result. Distinguish trading volume (the notional value or contracts traded) from open interest (outstanding positions). An API feed is not a probability oracle; it provides market observations that must be normalized, deduplicated, and analyzed with careful timestamping.

Designated Contract Markets Registry | CFTC Official register for checking the status of federally designated exchanges.
CFTC Orders Polymarket to Pay $1.4 Million Penalty | CFTC, January 3, 2022 Historical context on the difference between decentralized software and U.S. market registration.

Regulation, Incentives, and Market Integrity

U.S. regulatory framework. Event contracts that qualify as swaps or futures under the Commodity Exchange Act generally fall under the Commodity Futures Trading Commission (CFTC) and must be offered through an appropriate registered market structure if made available to the general U.S. public. A designated contract market (DCM) must satisfy core requirements concerning market integrity, surveillance, manipulation prevention, and other protections. Registration of an exchange should not be confused with government verification of every prediction or endorsement of every individual contract.

Sensitive event categories. Section 5c(c)(5)(C) of the Commodity Exchange Act addresses event contracts involving terrorism, assassination, war, gaming, or activity that is unlawful under federal or state law. The statute provides a process for determining whether relevant contracts are contrary to the public interest. These categories are especially relevant to the 2003 Pentagon controversy and the modern debate about contracts involving geopolitics, military operations, and professional sports.

Developments during 2024–2026:

  • 2024: litigation over Kalshi's election contracts led to a significant court ruling allowing those congressional-election markets to proceed, helping bring U.S. event trading to a wider public.
  • 2025: federally regulated sports-related event contracts expanded rapidly, bringing disputes with state gambling authorities and tribal gaming interests. Sports dominated many exchanges' trading volumes, which means total prediction-market volume cannot be treated as equivalent to independent economic or political forecasting activity.
  • March 2026: the CFTC issued an advance notice of proposed rulemaking asking how to apply market safeguards to prediction markets, including potential manipulation and the handling of inside information.
  • June 2026: the CFTC proposed a more detailed public-interest review framework for contracts in the law's enumerated sensitive categories. A proposed rule isn't the same as a final effective rule.
  • Throughout 2026: federal and state litigation continued over whether federally regulated sports contracts also fall under state gambling laws. Legal access and the availability of individual contract categories remained jurisdiction-dependent.
  • October 2026: the Pentagon's nonpublic-information directive highlighted the operational-security dimension that ordinary consumer-focused regulation may not fully address.

Specific integrity risks

  • Inside information and conflicts of interest: officials, speechwriters, campaign workers, athletes, employees, and contractors may learn information that others cannot lawfully obtain or may influence the outcome of an event.
  • Manipulation and wash trading: a participant may trade to create misleading prices or volumes, and thin markets can be distorted more readily than deep markets. Unusual trading is a reason for investigation, not automatic proof of fraud.
  • Moral hazard: a contract on an event someone can cause, prevent, or accelerate may create destructive incentives, particularly when violence or sabotage is involved.
  • Resolution risk: outcomes can be disputed; oracles, reporting agencies, and contract administrators can be mistaken or subject to pressure.
  • Consumer risk: an apparently straightforward Yes/No choice can conceal the possibility of losing the entire stake, rapid-fire speculative behavior, fees, and legal restrictions.
  • Forecast feedback: media coverage of a market price may itself change public behavior or the event being forecast, creating feedback loops in stock-market predictions and related phenomena in event markets.

A disciplined interpretation: prediction-market odds can be one input in a forecast, alongside underlying evidence and independent models. They should not substitute for a decision process, nor should a movement in prices be treated as confirmed intelligence. For national-security or other sensitive domains, access restrictions and approval processes may be more appropriate than public cash markets.

Prediction Markets: Policy Issues for Congress | Congressional Research Service, March 20, 2026 Legislative and jurisdictional background, including state gambling disputes.
CFTC Seeks Comment on Prediction Markets Rulemaking | CFTC, March 12, 2026 Official advance notice and scope of issues.
CFTC Seeks Comment on Event Contracts Involving Enumerated Activities | CFTC, June 10, 2026 Proposed review process for sensitive event categories.
CFTC Enforcement Division Issues Prediction Markets Advisory | CFTC, February 25, 2026 Official examples of improper trading and exchange enforcement.

AI Forecasting and Prediction-Market Agents

An AI forecasting system and an AI trading system should be evaluated separately. One asks how likely is this outcome? The other asks would a trade at today's executable price have positive risk-adjusted value after all costs? Confusing those questions is an easy way to build a persuasive-looking but unprofitable agent.

A useful architecture:

  1. Event ingestion: read the market catalog and rules; retrieve prices, official statistics, news, and primary-source documents with timestamps.
  2. Evidence review: identify independent signals and their provenance. Exclude information published after the forecast timestamp and evidence already embedded in the test outcome.
  3. Probability model: establish a baseline from historical base rates or an established forecast, then update with statistical models, language models, or a structured panel of independent forecasters.
  4. Cross-check: compare model probability with the executable bid/ask spread, related contracts, survey forecasts, and public data. Check whether logically related probabilities are coherent.
  5. Sizing and controls: require a minimum estimated edge net of fees, limits on exposure and drawdowns, restrictions on sensitive events, and human approval for any real-money trades.
  6. Post-event evaluation: archive the pre-event model probability, contemporaneous market probability, outcome, Brier score, realized return after costs, and a record of failed or canceled predictions.

Recent research: the 2025 ICLR ForecastBench project evaluates AI predictions on questions whose answers are not known when forecasts are submitted, reducing historical training-data leakage. In its reported comparison, expert human forecasters outperformed the strongest evaluated LLM. That result is a useful caution against treating plausible AI explanations as evidence of superior forecasting skill. In contrast, 2026 prediction-market microstructure research reports that prices can aggregate information efficiently in particular settings, though participation, resolution design, and liquidity remain important.

Recommended use: begin with a dashboard that collects and scores prices, then a research assistant that explains major odds changes with verifiable sources, then a paper-trading agent with explicit economic and ethical constraints. Autonomous real-money execution should come only after a sustained, forward-tested demonstration of calibration, positive net expected returns, and operational safeguards. Even then, market behavior can change as competitors learn from the same public evidence.

ForecastBench: A Dynamic Benchmark of AI Forecasting Capabilities | Karger and colleagues, ICLR 2025 Forward-looking evaluation of language-model forecasts compared with human forecasters.
SoK: Market Microstructure for Decentralized Prediction Markets | Rahman, Al-Chami and Clark, 2025 Technical survey of price formation, settlement, and manipulation-resistant market design.

Conversational AI is useful in trading, but mostly not in the way the "ask ChatGPT what to buy" videos suggest. Where the evidence is strongest:

  • Reading and summarizing: filings, earnings call transcripts, central bank statements, and news, faster than any human team. Research shows advanced LLMs extract real signal from headlines (see Recent Developments).
  • Research assistance and coding: writing data pipelines, backtests, and broker API code, provided a human checks the logic. LLM-written backtests frequently contain look-ahead bias (using future data) and survivorship bias (testing only stocks that still exist).
  • Structuring a process: checklists, trade journals, and risk rules that reduce impulsive decisions.

Where it is weak or risky:

  • Predicting prices from charts or raw price series. There is little evidence LLMs beat simple baselines here, and they are prone to confident, plausible-sounding pattern stories.
  • Remembering the future. Asking a model to "backtest" a strategy over a period included in its training data invites contamination: it may already know what happened.
  • Autonomous execution without guardrails. Live contests show large losses from overconfident position sizing. Use paper trading, hard position and loss limits, and human approval for real orders.

Prompting tips for trading research (see Prompt Engineering (PE)): give the model point-in-time data rather than asking it to recall facts; ask for a probability and the reasoning, not a buy or sell instruction; ask it to list what would make the view wrong; and log every prompt and response so you can evaluate the model's track record honestly later.

Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube

Build Trading Bot - with Alpaca Markets API

Alpaca is a brokerage that allows you to buy and sell stocks on US exchanges. The broker has gained popularity as it does not charge a commission for buying or selling stocks. What makes Alpaca unique, in addition to commission-free trading, is that it is very much focused on API trading. Today it also supports ETFs, options, and crypto through the same API, offers free paper trading accounts for testing, and provides an official MCP server so that AI assistants can research markets and place orders using natural language. Commission-free does not mean cost-free: you still pay the bid-ask spread, market impact, regulatory fees, and crypto trading fees.

Architecture of a responsible trading bot

Layer What it does Common mistakes
Data Streams prices, bars, and news (REST for history, WebSockets for live data) Using adjusted historical prices that leak future splits or dividends; ignoring delisted stocks
Signal Turns data into a forecast (rules, ML model, or LLM judgment) Look-ahead bias; testing on data the LLM has already seen
Risk and sizing Converts the forecast into a position with limits on size, exposure, and daily loss Sizing by conviction alone; no maximum drawdown stop
Execution Places orders with appropriate types (limit rather than market in thin stocks) Market orders in illiquid names; ignoring spread and slippage
Controls Kill switch, duplicate-order checks, rate limits, human approval for live orders Letting an agent place unlimited orders; storing API keys in code or prompts
Monitoring Logs every decision and compares live results with the backtest No record of why trades happened; not noticing when live diverges from test

A safe path from idea to live money: write the hypothesis down before testing; backtest with realistic costs and point-in-time data (see Backtesting); count how many variations you tried and discount the result accordingly; paper trade for weeks or months; then trade small with hard loss limits. If an AI agent can place orders, restrict its tools (Alpaca's MCP server supports toolset filtering), keep it on a paper account until it has a track record, and require confirmation for real trades.

Python: I Built a Trading Bot with ChatGPT | Siraj Raval

This early-2023 tutorial combines ChatGPT-generated code with FinRL, the open-source deep reinforcement learning library from the AI4Finance Foundation, and connects it to Alpaca for paper or live trading. It remains a useful illustration of the plumbing (data download, environment setup, agent training, order placement). Keep three caveats in mind: Python 3.7 reached end of life in June 2023, so expect to update dependencies to a supported version; reinforcement learning agents trained on a few years of daily prices are highly prone to overfitting, so treat any backtest curve as a starting hypothesis; and Siraj Raval's work drew public criticism in 2019 over plagiarism and a paid course, so verify code and claims independently. FinRL itself is actively maintained, runs annual open contests, and has spawned related projects such as FinGPT for financial language models.

Load video
YouTube

JavaScript: Build a ChatGPT Trading Bot With Real Time News | Cooper Codes

WebSockets to connect to a live news event data source from Alpaca Markets, analyze the headlines of live news, and then make decisions to buy/sell stocks based off of ChatGPT's reaction to the breaking news. A great introduction to WebSockets, API trading, and ChatGPT API calls.

WebSockets is a next-generation bidirectional communication technology for web applications which operates over a single socket and is exposed via a JavaScript interface in HTML 5 compliant browsers.

This design is a small-scale version of the strategy Lopez-Lira and Tang studied academically, so their findings apply directly. The model may read the headline correctly, but the initial price move usually happens before an LLM API call returns, so a bot like this is effectively trading the slower drift that follows. That drift is strongest in small stocks, where spreads are wide, and it has been shrinking as more traders use LLMs. Improvements worth adding: score only headlines about a single company, ignore duplicates and stale stories, use limit orders, cap position size and daily loss, and log every headline, model response, and fill so you can measure whether the bot actually has an edge.

Load video
YouTube

Other Videos

Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube

Other Siraj Raval Videos

Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube

Further Viewing and Sources

Key Sources

Machine learning and return prediction

Overfitting, backtesting, and research discipline

LLM agents and live trading

Market structure, stability, and integrity

Regulation and investor protection

Additional Viewing

Load video
YouTube

Introduction to Machine Learning in Finance (John Hull, University of Toronto; guest lecture 23 in MIT 18.642 Topics in Mathematics with Applications in Finance, Fall 2024). The author of the standard derivatives textbook walks through supervised, unsupervised, and reinforcement learning with finance examples, including neural networks for option pricing and reinforcement learning hedging strategies that reduce transaction costs. It is a rigorous, hype-free grounding for the execution and risk sections of this page. Watch on YouTube

Load video
YouTube

MIT Economist Andrew W. Lo on Finance, AI, and Human Behavior (MIT OpenCourseWare, Chalk Radio, 2025). Lo, director of the MIT Laboratory for Financial Engineering, explains his Adaptive Markets Hypothesis, why human behavior makes markets sometimes efficient and sometimes not, and why LLMs are not yet ready to act as trustworthy financial advisors. It pairs well with the non-stationarity and reflexivity discussion above. Watch on YouTube

Load video
YouTube

Futarchy and Prediction Markets: Robin Hanson Interview (Hodl Up, September 19, 2026; 44 minutes). Hanson discusses market design, why informed trading is controversial, decision-conditional markets, the difficulty of creating useful forecasting contracts, and the history of the Pentagon's Policy Analysis Market. The interview provides a practitioner's perspective to compare with regulatory and academic research. Episode and video

Retrieved from "https://primo.ai/w/index.php?title=Market_Trading&oldid=51957"