Why our model doesn't beat Pinnacle — and why that's the point

Our model doesn't beat Pinnacle — nothing reliably does. It matches the sharpest line in the world, and that's the point: it's a calibrated fair line. The product is finding where every other book disagrees with it.

That's the whole thesis. This essay is the receipts.

What "beating Pinnacle" would mean

Pinnacle's closing line is the most accurate public forecast of a tennis match that exists. It's sharpened by the one mechanism that beats any model: people betting real money into it all day, with the book raising limits on winners instead of banning them. If you can consistently buy better prices than Pinnacle's close — positive closing-line value — you have a real edge. If you can't, you don't, whatever your recent results say.

So before selling anything, we asked the question the honest way: does our model, evaluated only on matches it had never seen, beat the vig-removed Pinnacle close?

Six pre-registered answers: no

We ran six studies, each pre-registered — thresholds fixed before looking at results, walk-forward only (a model never grades years it trained on). Every confidence interval against the Pinnacle close includes zero:

| Study | What it tested | ROI vs Pinnacle close (95% CI) | | --- | --- | --- | | Deployed-filter revalidation (2024 out-of-sample) | The live filter on corrected market features | −5.7% [−28.0%, +16.7%], N=93 | | Multi-year honest walk-forward (2019–24) | Per-year retrained models, pooled | +2.3% [−7.7%, +12.2%], N=474 | | Style-matchup features | Stopped at its own pre-registered kill-switch | no-go (underpowered) | | Schedule/physical features | One pre-registered spec, honest walk-forward | +1.5% [−8.7%, +11.7%]; incremental −2.4% | | Pressure features | Data source no longer exists | no-go | | Full methodology audit, threshold sweep | Deployed cell + every neighbour | +3.7% [−2.2%, +9.7%], N=1,537 — every Pinnacle-close CI includes zero |

We could have quoted the point estimates and skipped the brackets. Most people selling picks do. The brackets are the finding.

What we actually have: calibration

A model that can't beat Pinnacle can still do something valuable: agree with it independently. Our model is calibrated — when it says 60%, the thing happens about 60% of the time. Across 11,769 walk-forward matches the mean bias is +0.0007. We calibrate per surface with isotonic regression, because raw scores drift differently on clay, grass, and hard courts.

Calibration is checkable, and we publish the check: the Receipts page regenerates monthly from every settled prediction, buckets what we predicted against what happened, and shows the counts. It also shows the misses. Right now our 90–100% bucket is running overconfident — we said 94% on average and it happened 81% of the time (n=64). That's a small sample and it may wash out; if it doesn't, it's a model defect and it stays on the page either way. A calibration record that only ever looked perfect would be the thing to distrust.

Why does a calibrated fair line matter if it can't out-predict Pinnacle? Because most books aren't Pinnacle. A fair line is a measuring stick, and the measuring only gets interesting when someone else's number is wrong — which brings us to where tennis prices go soft, the validated half of this product.

The bugs we caught, because you should ask

Any pipeline that produces only good news is broken. Ours has produced spectacular numbers twice, and both were bugs we found, documented, and retracted:

• A derivatives backtest once showed +10.21% ROI — on odds the code had fabricated (a hardcoded 1.91 standing in for real prices). Against real Pinnacle totals lines the same strategy was −3.1%. Retracted; the rule "never fabricate, simulate, or interpolate odds" is now written into the project's constitution.
• A walk-forward once showed +19.78% — because a script loaded the production model and graded it on years it had trained on. The honest rerun said +5.7% with a CI touching zero. Retracted.

The discipline that catches these — pre-registration, adversarial diagnostics on any surprising positive, CIs always — is the same discipline behind every number on this site.

The standing offer

Everything above is falsifiable, and we've pre-committed to the consequences: if the public ledger's aggregate closing-line value on filtered picks is ≤ 0 over any rolling 6-month window with n ≥ 100 picks, the 5–12% edge claim comes off the site and the positioning reverts to pure process and education. That rule is printed on the Receipts page, in public, where we can't unprint it.