Walk-forward backtest results for the FightIQ V12 dual-ensemble model. Every prediction made using only data available before fight night — no hindsight, no cheating, no exceptions.
Walk-forward period: Jan 2020 — Apr 2026 • 1,990 fights with odds data
Two independent model stacks (4 algorithms each) make separate predictions. The final output blends both ensembles, reducing variance and overfitting risk.
Gradient-boosted trees, logistic regression, neural network, and random forest. Each brings different strengths — the ensemble captures what no single model can.
Every prediction in this backtest was made using only data available before that fight. The model is retrained on a rolling window — no future information ever leaks in.
Time-weighted training data with a 5-year half-life. Recent fights count more, but deep history still matters. The sport evolves — the model adapts.
The most common mistake in backtesting is data leakage — accidentally using future information to make predictions about the past. This makes results look incredible on paper but useless in reality.
Walk-forward testing eliminates this entirely. Imagine you're standing in January 2022 and want to predict a fight. The model can only use data from before that date — fighter records, historical stats, training camp info, all from the past. It makes a prediction, we record it, then move forward in time.
We repeat this for every fight from 2020 to 2026. The model is periodically retrained as new data becomes available, but it never sees the outcome of a fight before predicting it. This is exactly how the model runs in production today.
The 66.2% overall accuracy and 81.3% LOCK+HIGH accuracy you see on this page are real out-of-sample results — the same kind of performance you can expect going forward.
Check out the latest UFC predictions — every fight graded by confidence tier, with full probability breakdowns.