How to Predict Odds Movement: A Modelling Approach

"Predict odds movement" is not one problem. Before choosing features, pin down exactly what is being forecast.

Horizon. Movement from now to the close, or over the next 15 minutes, or until the next injury report. Each horizon has different drivers.

Units. Raw price changes are not comparable across favourites and longshots. Convert to no-vig implied probability and model the change in that, or in log odds. For spreads and totals, model the line and the price separately.

Book. Movement at a sharp, high-limit market and movement at a slower retail book are different targets. The second is often partly predictable from the first.

Form. Direction only (classification), size (regression), or a full distribution. A distribution is the honest choice, because most of the time the best forecast is a small move centred near zero.

A model is a structured guess about these drivers.

Information. Injury and availability news, confirmed line-ups, starting pitchers, goalies, weather, and travel. These shift the underlying probability.

Price discovery between books. Low-margin, high-limit markets tend to move first, and other books follow. This propagation is the most persistent structure in the data.

Money and liability. Books adjust as exposure builds on one side, weighted by who is betting rather than how much.

Scheduled stages. Limits rise and prices tighten as the start approaches; many markets move most in the final hours.

Model refreshes. A book's own pricing model updating can move a line with no external cause.

Steam versus news

Use only information available at the moment of prediction, with timestamps proving it.

• Gap to the reference market. The difference between a book's no-vig probability and the sharp or consensus price. Outliers tend to converge, and this is often the strongest single feature.
• Cross-book dispersion. Wide disagreement across books signals unsettled pricing.
• Recent velocity. Movement over the last few windows, at this book and at the reference.
• Time to start. Movement concentrates near the close; interact this with everything else.
• Book identity and market type. Some books lag systematically; props and smaller leagues behave differently from main markets.
• Key numbers. In NFL spreads, lines around 3 and 7 behave differently from lines between them.
• Scheduled news windows. Whether an injury report or line-up release is still pending.
• Limits, where available. Low limits mark prices the book itself expects to move.

Betting markets are reasonably efficient. A sharp price already reflects public information, so its future changes are driven largely by information not yet public, which by definition is not in your features. For a reference market, the sensible baseline forecast is no change, and beating it consistently is difficult.

What remains predictable is mostly structural: slower books converging to faster ones, and outliers reverting to consensus. Those effects are real but short-lived and small, and they are the same effects that books themselves watch and correct.

Data quality compounds the difficulty. Sparse polling hides intraday moves, collection timestamps can lag real change times, and forward-filled gaps fabricate periods of stability. A model trained on that data learns the collection process, not the market.

Compare against the right baselines. Always report the no-change forecast and a simple convergence rule (each book moves toward the reference). A model that does not beat these has not learned anything useful.

Split by time. Train on earlier periods and test on later ones. Random splits leak information across events on the same day.

Audit for leakage. Every feature must carry a timestamp earlier than the prediction time. Features built from collection-time data, closing-day aggregates, or post-hoc injury tags are the usual culprits.

Measure in probability units. Mean absolute error of the predicted change in implied probability, calibration of predicted move distributions, and directional accuracy conditioned on predicted size. Directional accuracy over all cases rewards predicting tiny moves.

Check whether the forecast is actionable. A move that is predictable only seconds before it happens, or only at a book with low limits, is a research finding rather than a usable signal.

Validate against the close. If predictions made earlier systematically point toward where the closing price ends up, the model is capturing information the market later priced in.

Collect per-book snapshots at a density matched to the horizon, plus a timestamped news log. Convert prices to no-vig probabilities. Build features strictly as of each prediction time. Fit a simple model first, such as a regularised linear model on gap-to-reference and time to start, then try gradient boosting only if it beats that on time-split tests. Report results by sport, market type, and time-to-start bucket, since averages hide where any signal lives.

This page describes data and modelling methods and is not betting advice.