Strategy graveyard #2: we could predict line moves — and still lost money

Second entry in our strategy graveyard. This one hurts more than most, because the strategy we're burying actually worked — by the metric almost everyone in the analytics community treats as the gold standard. It just didn't pay.

The idea: bet before the line moves

"Steam chasing" is betting in the direction the odds are about to move: take 2.20 now, watch it close at 2.05, pocket the difference as closing line value. If you can predict line movement, you should — in theory — print CLV, and CLV is supposed to be the one metric that cannot lie.

So we tested whether opening odds errors are predictable from public data (Elo ratings and form versus the opening line). The honest answer surprised us: yes. The correlation between our signal and the subsequent open-to-close move was +0.04 to +0.08 — small, but positive in 4 out of 4 out-of-sample seasons, with clean controls. The market's opening line really does contain a sliver of publicly recoverable error.

Most betting content would end here, sell you the model, and show you the CLV chart. Here is what actually happened when we followed the signal with simulated money.

Autopsy: three ways to die on a real signal

Death #1 — the margin ate the edge. Following the signal at Pinnacle's opening price captured +1.9% CLV. Pinnacle's margin on those markets: ~3.3%. The signal was real; it was also smaller than the toll booth. You predicted the market and still paid for the privilege.

Death #2 — the winner's curse. So we "improved" it: take the best available odds across all bookmakers instead. CLV jumped to +5.2% — and the simulated bankroll fell to −2.4% ROI. This is the winner's curse, and it is the most under-taught concept in betting: the maximum price for any outcome is, disproportionately often, quoted by whichever bookmaker has made a mistake against themselves that they will shortly fix — or by one whose price you can't actually get paid at size. Line-shopping against a prediction signal selects, systematically, the prices that flatter your CLV and starve your wallet. CLV measured against the best price anywhere is a vanity metric. Same-book CLV — entry and close at the same bookmaker — is the only honest version, and it's the one our track record uses.

Death #3 — the execution window. The mispricings that drove the signal live minutes, not hours. By the time a public-data signal fires and a human places the bet, the softest prices are gone. What remains is the average price — and the average price contains the margin. See death #1.

What survived

The signal itself goes in the "true but useless" drawer — alongside our finding that 67% accuracy loses money. Three portable lessons:

1. A real signal is not an edge. An edge is a signal bigger than the cost of acting on it. Margin, execution lag and the winner's curse are the three costs everyone forgets to subtract.
2. Audit the CLV benchmark. When a tipster shows you closing line value, ask: measured against which book? If the answer is "best price at the time" — the number is inflated by construction. Our tipster audits use same-book CLV for exactly this reason.
3. If line moves were exploitably predictable, the market would already have eaten it. The sliver we found is precisely the size the margin allows to survive. Markets are not perfectly efficient — they are efficient enough, which for your bankroll is the same thing.

All numbers from our model investigation — full context in the 50-models autopsy. Next in the graveyard: whichever strategy dies next in our pre-registered live tests.