ROI tells you how much a football prediction model returned on average; maximum drawdown tells you the worst stretch you would have had to sit through to get there. Two models can post identical returns and still be completely different propositions — one grinding steadily upward, the other losing a third of a simulated bankroll before recovering. If you are comparing prediction models on a benchmark, drawdown is the number that tells you what the ride actually felt like, and it is arguably more informative than the return itself.
Why drawdown complements ROI
ROI is an aggregate. It collapses an entire sequence of results into a single ratio, and in doing so it throws away the order in which things happened. Maximum drawdown keeps that order. It is defined as the largest peak-to-trough decline in a running equity curve over the evaluated period, usually expressed as a percentage of the peak, and it is a path statistic rather than a point statistic.
That distinction matters for three reasons.
First, sequencing is real. A model that goes 3–7 with its first ten selections and 7–3 with its next ten has the same 10–10 record as one that does the reverse, but the first model experiences a far deeper drawdown. Same ROI, different risk.
Second, drawdown is what stress-tests the assumption underneath a simulated bankroll. Any simulation has to assume a stake size — flat units, a percentage of bankroll, or something proportional to an estimated edge. Drawdown is the metric that shows whether that assumption survives a bad run of results without the simulated bankroll approaching zero.
Third, it separates models that look similar on a leaderboard. On AlphaBench, models such as deepseek-flash (DeepSeek), gpt-5-mini (OpenAI) and grok-build-0.1 (xAI) are evaluated on the same settled fixtures under the same scoring rules. When their ROI figures land close together, drawdown and the other risk-aware metrics are what break the tie. A model with a slightly lower return but a materially shallower drawdown is often the more stable performer over a long season.
Worth saying plainly: drawdown is a descriptive measurement of what a model's published selections would have produced. It is not a prediction of future losses, and nothing here is betting advice.
Pilot calculation
The calculation is mechanical. You need four things: every settled selection in chronological order, the price recorded at the time the selection was logged, the result of the fixture, and a staking rule.
The steps:
1. Order the selections by kick-off time. Sequence is the whole point; sorting by anything else destroys the metric. 2. Apply each result to the running bankroll. With flat staking of one unit at decimal odds o, a winner adds (o − 1) units and a loser subtracts 1 unit. 3. Record the running peak — the highest bankroll value seen so far. 4. Compute the current drawdown as (peak − current) ÷ peak. 5. Maximum drawdown is the largest value that expression ever reaches.
Here is an illustrative example. The numbers below are invented to demonstrate the arithmetic — they are not AlphaBench results for any model.
| # | Odds | Result | P/L (units) | Bankroll | Peak | Drawdown | |---|------|--------|-------------|----------|------|----------| | 1 | 2.00 | Win | +1.00 | 101.00 | 101.00 | 0.0% | | 2 | 1.80 | Loss | −1.00 | 100.00 | 101.00 | 1.0% | | 3 | 2.50 | Loss | −1.00 | 99.00 | 101.00 | 2.0% | | 4 | 2.20 | Loss | −1.00 | 98.00 | 101.00 | 3.0% | | 5 | 1.90 | Loss | −1.00 | 97.00 | 101.00 | 4.0% | | 6 | 2.40 | Win | +1.40 | 98.40 | 101.00 | 2.6% | | 7 | 3.00 | Loss | −1.00 | 97.40 | 101.00 | 3.6% | | 8 | 2.10 | Win | +1.10 | 98.50 | 101.00 | 2.5% | | 9 | 1.75 | Loss | −1.00 | 97.50 | 101.00 | 3.5% | | 10 | 2.60 | Win | +1.60 | 99.10 | 101.00 | 1.9% |
Starting bankroll: 100 units. Flat stake: 1 unit. Peak: 101.00 after selection 1. Trough: 97.00 after selection 5. Maximum drawdown: (101.00 − 97.00) ÷ 101.00 = 3.96%. Final bankroll after ten selections: 99.10 units, for an ROI of roughly −0.9% on ten units staked.
Two lessons fall out of this table. The first is that a 4-loss run in the middle of a small sample produced the entire drawdown — the wins afterwards did not erase the fact that the sequence dipped below 97 units. The second is that the staking rule scales the result. Run the identical selections at a flat 10-unit stake and the peak becomes 110 units, the trough 70 units, and maximum drawdown jumps to roughly 36%. Same model, same predictions, same ROI — a completely different risk profile.
In practice, a benchmark would compute this over an entire season of settled fixtures. Every match adds a step to the curve; a fixture such as Khovd Western versus Deren in the Premier League is just one row in a long sequence. AlphaBench publishes settled fixtures precisely so that these sequences can be reconstructed and audited rather than taken on trust.
Using drawdown with CLV and calibration
Drawdown on its own is easy to misread, because a shallow drawdown can simply mean the sample was short or the sequence was lucky. It is most useful alongside the metrics that describe why a model performs the way it does.
- Brier score and log loss measure the quality of the probabilities themselves. They are proper scoring rules: they reward honest, well-shaped probability estimates rather than merely correct picks. A model can have a good Brier score and still show a deep drawdown if its edge is thin and its staking is aggressive.
- Calibration checks whether predicted probabilities match observed frequencies. A model that says 30% should see roughly 30% of those selections land. Poor calibration is a common source of drawdowns that look inexplicable from the win-loss record alone.
- CLV (closing line value) measures whether the price recorded when a selection was logged was better than the closing price. It is the most process-oriented of the group, and it is the one that best distinguishes skill from a favourable run of results. A model with positive CLV and a deep drawdown is telling you something different from a model with negative CLV and a deep drawdown.
- ROI remains the headline number, but it should be read as a companion to drawdown, never as a substitute.
The useful pattern is to look at these together. Deep drawdown plus poor calibration suggests the probability estimates are the problem. Deep drawdown plus good calibration and positive CLV suggests the model may be sound but the sample is small, the variance is high, or the assumed staking is too aggressive for the edge. Shallow drawdown plus a strong Brier score and positive CLV is the combination most people are looking for — and it is the combination that most deserves scepticism until the sample is large.
Sample size deserves emphasis here. Maximum drawdown is an extreme-value statistic, so it is noisy on short histories and tends to grow as more fixtures are added. It should always be reported with the number of selections and the time span it covers, or it means very little. Our methodology page sets out how AlphaBench handles evaluation windows, and how AlphaBench scores models explains how the individual metrics fit together.
Responsible use
A few guardrails keep this metric honest.
It is a historical simulation, not a forecast. Maximum drawdown describes a sequence that already happened. The worst stretch in the next hundred fixtures is unknown and could easily be deeper.
It is an assumption-dependent number. Change the stake size, change the staking rule, or change the starting bankroll, and the drawdown changes. Any reported figure is only meaningful with its staking assumption stated alongside it. This article describes how simulated bankrolls are modelled for comparison purposes; it is not betting advice, and nothing here should be read as a recommendation about staking real money.
It is vulnerable to the usual biases. Lookahead bias, survivorship in the fixture sample, using prices that were not actually available at the time, and quietly dropping selections that turned out badly will all flatter the number. A trustworthy benchmark logs selections before kick-off and settles them afterwards, whether they win or lose.
A deep drawdown is not automatically disqualifying, and a shallow one is not automatically reassuring. Drawdown is one lens. It is most valuable when it is read next to ROI, calibration, Brier score, log loss and CLV — and when it is tracked over a long, boring, unedited run of fixtures rather than a hand-picked streak.
Related reading
- Methodology — how AlphaBench collects selections, prices and settled results.
- Brier score for football predictions — why proper scoring rules matter for probability quality.
- How AlphaBench scores models — how accuracy, calibration, CLV and risk metrics combine.
- Leaderboard — compare models such as deepseek-flash, gpt-5-mini and grok-build-0.1 on the same settled fixtures.