Home › Best Bets
Best Bets — every market, ranked by chance, gated on price
For each game, every market we can model, ordered by how likely it is to land. A selection reaches the board only when the price on offer pays more than that chance requires. Backtested walk-forward across four codes, on real closing prices where they exist.
How a selection gets on the board
- Price every market. An Elo rating per team, updated after every game, plus each side's recent scoring and conceding rate. That gives an expected margin and an expected total, turned into a probability through a normal curve whose width is fitted on earlier seasons only.
- Rank by chance. Highest probability first, across head to head, the line, the total and the period totals — the whole slate in one list.
- Throw away anything the price does not pay for. A 92% chance needs 1.09 to break even. If the market offers 1.05, it is cut, however confident the model is. This is the step that stops the board filling with favourites, and it is why the board is short some weeks.
- Grade it walk-forward. Every rating used to price a game comes from games played before it. Nothing is fitted on a result it then scores.
What the backtest found
Every selection the gate would have published, replayed across every season held for each code. The comparison is not 50% and not a coin flip: it is what the prices themselves implied those same selections would do. Beating that is the only thing that counts, because it is the only thing you are actually betting against.
NRL
Over 1,097 selections the board hit 54.6% against the 54% the prices implied — a gap of +0.6 points — and returned +2.3% on turnover at flat stakes. The price already contains what the model knows.
| Market | Selections | Hit | Implied | Edge | Return | Chance |
|---|---|---|---|---|---|---|
| Head to head | 694 | 56.6% | 56.2% | +0.4 | +2.7% | 83.2% |
| The line | 209 | 49.3% | 50.7% | -1.4 | -3.1% | 72.9% |
| Total points | 238 | 50.8% | 50.2% | +0.6 | +1.4% | 90.6% |
| Period totalsassumed price | 256 | 50.8% | 52.4% | -1.6 | -3.1% | 65.2% |
| Model chance | Selections | Hit | Implied | Edge | Return | Chance |
|---|---|---|---|---|---|---|
| 50-60% | 759 | 49.3% | 48.3% | +1 | +2.9% | 61.4% |
| 60-70% | 324 | 50.6% | 53.1% | -2.5 | -3.4% | 40.4% |
| 70-80% | 179 | 62% | 62% | 0 | -0.3% | 100% |
| 80-90% | 102 | 71.6% | 71.7% | -0.1 | -0.9% | 100% |
| 90-100% | 33 | too few to measure | ||||
AFL
Over 1,239 selections the board hit 52.5% against the 55.1% the prices implied — a gap of -2.6 points — and returned -4.7% on turnover at flat stakes. The price already contains what the model knows.
| Market | Selections | Hit | Implied | Edge | Return | Chance |
|---|---|---|---|---|---|---|
| Head to head | 790 | 54.7% | 57.5% | -2.8 | -5.1% | 11.9% |
| The line | 153 | 47.7% | 50.9% | -3.2 | -6.7% | 47.3% |
| Total points | 296 | 49.3% | 50.8% | -1.5 | -2.4% | 65% |
| Period totalsassumed price | 1,297 | 49.3% | 52.4% | -3.1 | -5.8% | 3.1% |
| Model chance | Selections | Hit | Implied | Edge | Return | Chance |
|---|---|---|---|---|---|---|
| 50-60% | 1,604 | 48.7% | 50.5% | -1.8 | -3.3% | 14.3% |
| 60-70% | 510 | 46.3% | 53.1% | -6.8 | -13.2% | 0.2% |
| 70-80% | 239 | 59.8% | 61.4% | -1.6 | -4.2% | 65.5% |
| 80-90% | 147 | 68.7% | 71% | -2.3 | -3.1% | 59.2% |
| 90-100% | 36 | too few to measure | ||||
NFL
Over 1,422 selections the board hit 49.5% against the 52.4% the prices implied — a gap of -2.9 points — and returned -5.5% on turnover at flat stakes. The price already contains what the model knows.
| Market | Selections | Hit | Implied | Edge | Return | Chance |
|---|---|---|---|---|---|---|
| The lineassumed price | 129 | 51.2% | 52.4% | -1.2 | -2.3% | 84.9% |
| Total pointsassumed price | 297 | 50.8% | 52.4% | -1.6 | -2.9% | 63.6% |
| Period totalsassumed price | 996 | 48.9% | 52.4% | -3.5 | -6.7% | 3% |
| Model chance | Selections | Hit | Implied | Edge | Return | Chance |
|---|---|---|---|---|---|---|
| 50-60% | 1,148 | 50% | 52.4% | -2.4 | -4.5% | 11.3% |
| 60-70% | 229 | 47.6% | 52.4% | -4.8 | -9.1% | 16.7% |
| 70-80% | 41 | too few to measure | ||||
| 80-90% | 4 | too few to measure | ||||
NBA
Over 1,683 selections the board hit 51.7% against the 52.4% the prices implied — a gap of -0.7 points — and returned -1.3% on turnover at flat stakes. The price already contains what the model knows.
| Market | Selections | Hit | Implied | Edge | Return | Chance |
|---|---|---|---|---|---|---|
| The lineassumed price | 767 | 55% | 52.4% | +2.6 | +5% | 15.3% |
| Total pointsassumed price | 140 | 50% | 52.4% | -2.4 | -4.5% | 63.1% |
| Period totalsassumed price | 776 | 48.7% | 52.4% | -3.7 | -7% | 4.4% |
| Model chance | Selections | Hit | Implied | Edge | Return | Chance |
|---|---|---|---|---|---|---|
| 50-60% | 1,300 | 52.2% | 52.4% | -0.2 | -0.3% | 93.6% |
| 60-70% | 317 | 48.6% | 52.4% | -3.8 | -7.3% | 19.4% |
| 70-80% | 49 | too few to measure | ||||
| 80-90% | 16 | too few to measure | ||||
| 90-100% | 1 | too few to measure | ||||
Is a stated chance honest?
A board ranked by confidence is only worth reading if the confidence means something. The first version of this model was badly overconfident: it called NBA selections 90% that landed 51% of the time, and scored worse than a predictor that simply says “50%” to every game. Every probability is now passed through a calibration step fitted on earlier games only and refitted as the seasons run, so a stated 70% is measured against what 70% actually did.
| Code | Worst gap before | Worst gap after | Brier before | Brier after | Coin flip | Better than a coin flip? |
|---|---|---|---|---|---|---|
| NRL | 14.3 pts | 5.4 pts | 0.249 | 0.241 | 0.25 | yes |
| AFL | 13.9 pts | 5.3 pts | 0.247 | 0.237 | 0.25 | yes |
| NFL | 13.1 pts | 7.3 pts | 0.253 | 0.243 | 0.25 | yes |
| NBA | 21.9 pts | 2.8 pts | 0.258 | 0.238 | 0.25 | yes |
Calibration makes the number honest. It does not create an edge: a perfectly calibrated model that agrees with the market still loses the margin. It is what makes everything below measurable.
Why the confidence bands matter more than the headline
If the model's confidence meant anything, the 90–100% rows would return more than the 50–60% rows. Read down the right-hand column of each band table and judge for yourself whether they do. That comparison is the honest test of whether "cream of the crop" is a real category or just the shortest prices on the slate.
Where the prices come from
The NRL and AFL carry real closing prices for head to head, the line and the total, so those returns are money. The NFL and NBA histories carry a closing line and total but no price, so standard −110 is assumed and the rows are marked. Period totals have no posted price in any code — no bookmaker publishes historical quarter lines — so those are graded against a fair benchmark and marked too. A return computed from an assumed price is not a return. The real-price figure is reported separately for every code that has one.
More
Built 2026-10-16. Break-even at standard −110 pricing is 52.4%, not 50%.