TL;DR — Order-flow imbalance predicts short-horizon direction with real statistical significance. It also loses money after spread costs: median net −6,199 bps. Both facts, same study, both published.
The setup
A full-day backtest across 6,307 active symbols (65,536 allocated), replaying the tape through the engine and evaluating an order-flow-imbalance signal at multiple thresholds. At extreme imbalance (32σ), AAPL showed a 59.48% hit-rate; across 5.9M samples the aggregate directional accuracy held at 55.28% — far outside what chance produces at that sample size.
The catch
Direction isn't profit. Every entry crosses a spread, and short-horizon edges are measured in fractions of it. Net of that cost, the median outcome is −6,199 bps. The signal is real. The trade isn't — at least not naively, not at these horizons, not paying the spread every time.
Why publish a losing backtest
Because the alternative is the industry default: show the hit-rate, hide the cost model, let the reader assume. A reviewer who sees you publish the loss knows they can trust your wins. That trade — credibility for vanity — is the cheapest positive expectancy on this entire site.