How to Use Strategy Lab

Strategy Lab asks one narrow question, over and over: if you had bet every time a particular pattern appeared in the odds, would you have made money?

That is all it measures. It is not a prediction engine and it is not our models — those are graded separately on the Performance page. Each strategy here is a fixed rule about how prices moved in the last hour before kickoff, applied to every match we recorded movement for, and settled at the price that was actually available.

Nothing is re-fitted after the fact. The rules were written down first, and the numbers are whatever they turned out to be — including the ones that lose money.

What a strategy is#

Every strategy is a short list of numeric conditions. For example, Weak Drop LOW → Under fires when all three of these are true:

• the league's historical Over rate is below 48% (a low-scoring league),
• the home price drifted down by between 0.05 and 0.09,
• the away price moved up by more than 0.03.

When all three hold, the strategy backs Under 2.5. Nothing else about the match is considered — not the teams, not the form, not our model's opinion.

That narrowness is the point. A rule with three conditions and no discretion can be tested honestly; a rule that you adjust as you go cannot.

Reading a strategy card#

Each card carries four numbers and two pieces of context.

Win rate, with a ±. The share of settled signals that won, followed by a 95% confidence interval. Read the interval, not the headline. A strategy showing 50.0% ±31pp from six signals is telling you that the true rate is somewhere between about 19% and 81% — which is to say, unknown. One showing 64.1% ±10pp from seventy-eight is making a much narrower claim. See sample size for why the number of bets matters more than the number itself.

ROI. Profit per unit staked at flat 1u stakes, using the real closing odds each signal was graded at. Positive is green, negative is red — and on this page, several are red.

Record and average odds. Wins, losses, and the average price. A strategy winning 45% of its bets at average odds of 1.89 is roughly breaking even before the margin; the same win rate at 2.20 would be profitable. Win rate alone never tells you enough.

Evidence tier: Early, Building, Tracked. This is the one most likely to be misread, so it is worth being blunt about:

The tier describes how much evidence exists, not whether the strategy works. It is derived from the signal count alone — Early is under 20 signals, Building is under 50, Tracked is 50 or more. A strategy earns "Tracked" by firing often, whether it made money or lost it.

At the time of writing one Tracked strategy is down roughly 14% over a hundred signals, while the best performer on the page is still labelled Early. The colour of the tier never means the strategy is good. That is what the ROI figure is for, and green and red are reserved for it.

The sparkline. A miniature of the cumulative profit curve. It answers a question the ROI figure cannot: was this a steady grind, or one lucky run that has been drifting ever since?

The equity curve and the signal log#

Open a strategy and you get its full history.

The equity curve plots cumulative profit at 1u flat stakes, oldest signal first. A curve that climbs steadily is a different animal from one that made everything in a three-week burst and has gone sideways since, even when both end at the same ROI.

The signal log lists every firing: the match, the score, the price it was graded at, and whether it won. Click any row to open that match in full. Click the chevron at the end of a row to see why that match was selected — each of the strategy's conditions, the value this particular match had, and whether it passed.

That breakdown is worth using, because the SIGNAL column is necessarily a summary. On the Weak Drop strategies it shows the away movement, while the condition that defines the strategy is the home drift. The expanded view shows both, with the thresholds beside them, so you can see the whole rule rather than one number from it.

One detail that trips people up: the movement metrics are calculated as opening price minus closing price. A positive value means the price shortened. A negative value means it drifted out. So "away odds move below −0.03" is asking for the away price to have risen.

Tuning a strategy#

Every threshold in a strategy is adjustable. Open Tune this strategy under the equity curve, move the sliders, and the whole record recomputes — signals, win rate, ROI, and the hold-out comparison — as fast as you can drag.

It runs entirely in your browser. The candidate matches download once and every recalculation happens locally, so tuning costs nothing and you can explore freely. Your settings are yours: they never change the published strategy or anyone else's view of it, and Reset restores the original rule.

The part that matters#

Here is the trap, stated plainly, because the tool makes it easy to fall into:

If you move sliders until the ROI looks good, you have not found an edge. You have fitted a rule to noise.

With enough thresholds and enough attempts, you can make almost any dataset look profitable in hindsight. The result means nothing, because you chose the rule after seeing the outcomes it would produce.

This is why the panel always shows two figures side by side:

• Fitted on the first 70% — your thresholds applied to the earlier part of the history.
• Held out — never fitted — the same thresholds scored on the later months, which played no part in your choice.

The split is chronological, never random. A random split would scatter matches from the same weeks across both halves and let future information leak into the fitted period, which flatters the result — the same class of mistake that once made a set of backfilled predictions look profitable when they were not.

When a tuned strategy earns well on the fitted half and loses on the held-out half, the panel says so directly. Believe it. That pattern is the signature of a curve-fit, and it is the most common outcome of enthusiastic tuning.

Also watch the sample size as you tighten a rule. Narrowing a threshold usually improves the headline ROI and shrinks the number of signals — and the confidence interval widens accordingly. A rule that returns +40% on nine bets has told you almost nothing.

What this page will not tell you#

Strategy Lab measures whether a market pattern paid, on the history we have recorded. It does not tell you that it will keep paying. Market inefficiencies get priced out; some of these patterns may already be gone.

The samples here are small, the market is genuinely difficult to beat, and several of these strategies lose money over their recorded history. We publish those alongside the winners deliberately — a page that only showed the profitable rules would be a different and much less useful thing.

Treat it as a research tool, not a tipping service. If you want to test whether a rule has real predictive value, the honest measurement is closing line value: whether your bets beat the price that was available at kickoff. ROI over a few dozen bets mostly measures luck.