...predict values

↺ Start over · Full treeYou are here: AI Solver › Predict and analyze › ...predict values

AI Solver map

Your current page is in bold 0 · Start Do I need AI? Full decision tree 1 · Predict & analyze tables, numbers, events Predict values Predict categories Forecast time series Cluster / segment Find outliers Make recommendations Find key drivers 2 · Understand unstructured data text, documents, images, audio Understand language Images & video Speech & audio 3 · Generate & converse Generate content Answer questions from my data Build an agent 4 · Decide & act Optimize decisions 5 · Build it right applies to every path Choose a model Customize a model Deploy AI Evaluate & trust AI New architectures Products & platforms

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

• AI Solver ... Regression ... ...predict categories ... ...forecast time series ... Evaluating Machine Learning Models ... Train, Validate, and Test

Use this page when the answer is a number: a price, a quantity, a duration, a risk score. This is supervised learning (regression): the model learns from rows where the right number is known. If the number is the next value of a series over time, go to ...forecast time series.

Start here

Answer top-down and take the first branch that fits; then compare two or three candidates on held-out data.

• Is your input mainly text, images or audio rather than a table?
  • Yes → turn the inputs into embeddings with a pretrained model and fit a regressor on top, or fine-tune a Deep Neural Network (DNN). See ...understand language and ...understand images and video.
• Is the target a special kind of number?
  • Event counts (calls per hour, claims per policy) → Poisson Regression; log-linear (or a Poisson/Tweedie loss in gradient boosting)
  • Rank-ordered categories (1–5 stars) → Ordinal Regression
  • You need a range, not a single point ("90% of deliveries arrive within ...") → quantile regression (Fast Forest Quantile Regression, or quantile loss in LightGBM/CatBoost), or wrap any model with conformal prediction for calibrated intervals
  • Big outliers should not dominate → robust loss (Huber), or Support Vector Regression (SVR), where small deviations are not penalized
• Must the model be simple to explain or audit?
  • Yes → Linear Regression (also your baseline for every project); Ridge Regression when features are correlated or variance is large; Lasso to drop useless features; Bayesian Linear Regression for small data with uncertainty estimates
• How many labeled rows do you have?
  • Small to medium (hundreds to ~10,000+ rows) → try a tabular foundation model (TabPFN, TabICL): no tuning, predictions in one pass, often better than tuned boosting on small tables
  • Medium to large → gradient-boosted trees (XGBoost, LightGBM, CatBoost; see Gradient Boosting Machine (GBM) and (Boosted) Decision Tree). Fast, accurate, handles missing values; CatBoost is strong when many columns are categorical
  • Very large, or mixed with unstructured inputs → Deep Neural Network (DNN), if you can accept long training times
• No data-science team, or need a strong result fast? → AutoML (AutoGluon, or the AutoML in your cloud ML platform) trains and ensembles many models for you

Quick picks

If you need...	Try first	Then compare
A fast, explainable baseline	Linear Regression / Ridge Regression	Lasso, Bayesian Linear Regression
Best accuracy on a table	LightGBM or CatBoost	XGBoost, AutoGluon ensemble, TabPFN on smaller tables
Accuracy with little data	TabPFN or TabICL	Bayesian linear regression, heavily regularized boosting
A prediction range	Quantile gradient boosting	Conformal prediction around any model, Fast Forest Quantile Regression
Counts	Poisson or Tweedie gradient boosting	Poisson Regression
Low memory, fast training	Decision Forest Regression	LightGBM
Text or image inputs	Embeddings + gradient boosting	Fine-tuned neural network

Classic options that still work but are rarely the first pick today: General Regression Neural Network (GRNN), Decision Forest Regression, Fast Forest Quantile Regression.

Real-world examples

• Arrival-time prediction — Uber starts from its routing engine's estimate and uses a deep learning model (DeepETA) to predict the remaining error, improving accuracy at very low latency. DeepETA: How Uber Predicts Arrival Times | Uber
• Cautionary tale: home prices — in 2021 Zillow shut down its home-buying business, citing the unpredictability of forecasting home prices, after a write-down of roughly $304 million on homes bought above its later price estimates. Accurate on average is not enough when the market shifts and every error costs real money. Zillow shuts down its home-flipping operation | Bisnow

What's new

• Tabular foundation models learn "in context" from your rows instead of being trained per dataset. TabPFN (published in Nature in 2025) beat tuned baselines on datasets up to 10,000 samples in seconds; newer versions and rivals (TabICL, TabFM) extend this. On large tables, gradient-boosted trees stay more cost-efficient. What are tabular foundation models? | DataCamp
• LLMs are not the default for predicting numbers from tables. Use them to turn free text into features, or to explain the model's results to users.

Watch out

• Leakage: a feature that is not known at prediction time (e.g., "final invoice amount") makes the model look great in testing and fail in use.
• Shift: monitor error over time and retrain; stress-test with past shocks.
• Report error in business units (minutes, dollars), not only R².

Measure success

MAE (average miss, in the target's units), RMSE (punishes big misses), MAPE or WAPE (relative error), R², and interval coverage when you predict ranges. Always compare against a naive baseline such as the mean or last known value. More: ...evaluate and trust AI

Learn more

• Choosing the right estimator | scikit-learn
• Accurate predictions on small data with a tabular foundation model (TabPFN) | Nature
• Why do tree-based models still outperform deep learning on tabular data? | L. Grinsztajn, E. Oyallon & G. Varoquaux
• Machine Learning: Trying to predict a numerical value | Stacey Ronaghan - Medium
• Evaluating Machine Learning Models
  • Feature Engineering and Selection: A Practical Approach for Predictive Models - 6 Engineering Numeric Predictors | Max Kuhn and Kjell Johnson
  • Regression Model Insights | Amazon Web Services
• Excel ... Documents ... Database; Vector & Relational ... Graph ... LlamaIndex

When a value is being predicted, as with stock prices, supervised learning is called regression. Prediction problems (e.g. What will the opening price be for shares tomorrow?) are a subset of regression problems for time series data. Machine learning algorithms explained | Martin Heller - InfoWorld

---

Where next? ◀ Predict & analyze · ...predict categories · ...forecast time series · ...find key drivers · ▶ ...evaluate and trust AI