Loading live data…•Loading live data…•Loading live data…•

Backtesting

Test it
on the past.

A hunch is cheap. Run your strategy against seasons of real historical odds — walk-forward by season, the vig included, the drawdowns shown — and see whether it actually held up, or just felt like it did.

REPLAYING 5 SEASONS
5
seasons of odds
Real
closing prices
Flat 1u
every bet
Walk-forward
no leakage

How it runs

Replay every game.
Grade every bet.

Pick a model, a market, and a stretch of seasons. Blitzen replays every game in that window, places the bet the model would have placed at the real odds, and scores it against what actually happened.

01

Set the strategy

Choose a model and market, a span of seasons, and your filters — a minimum model-vs-market gap, confidence tier, even the juice you actually pay.

02

Replay the seasons

A fresh model is trained per season on prior years only, then every game in the test season is bet at its real closing odds.

03

Grade the ledger

Each bet is scored win/loss/push against the real result — and against the close — rolling up into ROI, units, drawdown, and CLV.

No leakage

Trained on the past.
Tested on the unseen.

The fastest way to fake a great backtest is to let the model peek at the answer. We don't. Blitzen trains a fresh model for each season using only the seasons before it, then predicts that held-out season cold — every graded game is one the model never saw.

The window expands as it walks forward in time, exactly the way you'd actually bet: with the past, not the future. No leakage, no cherry-picked sample.

20222023202420252026
Predict 2023
Train
Test
Predict 2024
Train
Train
Test
Predict 2025
Train
Train
Train
Test
Predict 2026
Train
Train
Train
Train
Test
Train — only the pastTest — held-out, never seen

The whole story

Wins and losses.

A backtest that only shows green is a sales pitch, not a tool. Yours reports the full ledger: record, ROI, net units, the worst drawdown, the longest losing streak — at flat 1-unit stakes and real closing prices, so the vig is already working against you.

Break it down by season and by confidence tier, and run the same seasons on another market to compare. Then judge it yourself, with the losses in plain sight.

Graded against the close

Win rate isn't enough.

Where the line history allows, a backtest also reports a closing-line read: for the side each bet took, did the consensus line move toward it between an earlier pre-game price and the close? A winning streak can be luck; this is a second read on the same bets. It's a research measure, not a promise of profit. See how we measure it →

Your controls

Tune the test.

A backtest is only as honest as its assumptions, so the assumptions are yours to set.

Seasons

Any span from 2022 through the current season — one year or all of them.

Minimum gap

Only count a bet when the model’s number differs from the market’s by at least the gap you set.

Confidence tier

Strong-only by default; opt into leans and toss-ups to see the full distribution.

The juice

Set the spread/totals price you actually pay; moneylines use the real odds per game.

Don't trust it. Test it.

Backtest your model against past seasons of real lines. Free to start — bring your own read.