The promo is the scarce resource, not money and not candidates — without one every band is negative. Record what you hold and the allocator spends them on the races worth the most.

This is not arbitrage, and it is not a betting edge. The raw bets LOSE 8.2% of turnover — odds taken sit about 9% below Betfair SP and only 13% of bets beat market price. Every dollar of profit comes from bonus-back promotions, which add 21.6%. Net +12.8%.

Anything above 6.0 loses even with the promo attached. An odds ceiling of 6.0 is better in 4 of 5 years and keeps 91% of bets: ROI +13.4% → +14.8%, losing weeks 30% → 25%.

Profit is capped by promo limits, not bankroll — roughly $50 a bet, ~40 bets a week. $10,000 earns the same as $5,000 because the extra capital cannot be deployed. The measurable risk is not variance, it is restriction: bets beating Betfair SP fell from 22.5% in 2020 to 9–13% by 2023, while the promo return itself held at ~21.5% every single year.

Every backtest figure in this project inherits whatever yfinance got wrong, and until now that has been an assumption rather than a measurement. This pulls the same window from both sources and differences them. If they agree, the coverage matrix stands and IBKR buys us only bid/ask and realistic fills. If they diverge, the matrix needs rebuilding on better bars.

Every pattern against every market, 5 years of daily bars, 8 concurrent positions, $25,000, run twice — once gross and once with the real fee schedule. The gap between the two is the whole story. Cells are coloured by net profit factor against a break-even midpoint; each prints its own number, so the colour only reinforces.

Every constant and variable the strategy actually reads, where it lives in the code, and whether its value has ever been checked. Sweeps are one-at-a-time with everything else held at its deployed value — a full grid over eight parameters is about a hundred thousand runs and would surface a "best" combination almost entirely curve-fit to these five years. The narrower question is the useful one: is the deployed value defensible, or did it simply never get looked at?

Live results sit above the backtest in every comparison. Not one is statistically distinguishable from noise — confirming the VWAP gap needs about 264 closed trades and there are 22; the EMA_PB gap needs 96 and there are 3.