Method
Value betting vs tipping
On 31 of our 73 logged bets, the bet was on a different outcome than the tip. That is not an inconsistency. It is the difference between the most likely result and the best priced one.
Two different questions
The tip answers: which outcome is most likely? It is the highest of the three model probabilities, and it ignores the odds entirely.
The bet answers: which outcome, at the price on offer, has the highest expected value? That depends on both the probability and the price, and the two answers often differ.
A worked example from the log. The model made a home win the most likely
outcome and tipped it. The away side was priced at 5.60. The
model gave the away win enough probability that 0.21 × 5.60 = 1.18
— an expected return of 1.18 units per unit staked, against less than 1.00 for
the favourite. The bet went on the away side. The tip lost. The bet returned
+17.11.
How expected value is computed
For each outcome: EV = model_probability × best_available_odds − 1.
Positive means the price is longer than our probability justifies. The bet goes
on whichever of the three has the highest EV, and only if that EV is positive.
The model's probability is compared against the market's margin-free probability, not the raw odds. Bookmakers build in an overround; stripping it out is the difference between measuring a real edge and measuring their commission.
Staking: quarter-Kelly, capped
Stake size comes from the Kelly criterion, which sizes a bet by edge divided by odds. Full Kelly is famously volatile, so we use quarter-Kelly and additionally cap any single bet at 5% of bankroll.
No wager is ever actually placed. The P&L on the record page is paper arithmetic on a flat notional bankroll, and we say so rather than implying a trading history.
The refusal rule
When the model claims an edge larger than 12 percentage points against a market priced by ten or more bookmakers, we do not bet it. It is flagged SKIP instead.
The reasoning is unglamorous: an edge that large against a deep market is almost always our error, not their mispricing. It usually means thin data on a newly promoted side or a bad fit early in a season. A model that bets its biggest apparent edges is a model that bets its biggest mistakes.
Seeing zero value bets on a given day is the system working, not a fault.
What this costs
Value betting has a lower strike rate than tipping, by construction. Longer prices win less often. Ours is 19 of 73, or 26.0%, against a tip accuracy of 43.3%. The theory says the longer prices should more than compensate over enough bets.
They have not, so far: −46.02 and −21.4% ROI. Seventy-three bets is a small sample for a strategy whose whole premise is variance, but we are not going to call a losing number a win because the method is sound in principle.
Reading it on the board
The Tip column is the most likely outcome. The Bet column is the value pick with its price, coloured by whether the bet won rather than the tip — green won, red lost. They disagree often, and now you know why. Full column guide.
Figures current as of 27 August 2026. Parameters: quarter-Kelly, 5% maximum stake, 12-percentage-point sanity ceiling on edge.