Statistical Edge and Signal-Based Strategies
A signal is anything that predicts future price movement better than chance. Modern trading firms run many of them, combined into quoting decisions.
Signals that actually work at short horizons
Order book imbalance. Far more size bid than offered predicts short-term upward movement. One of the most robust known signals, and one of the most heavily used.
Trade flow imbalance. Persistent aggressive buying predicts continued upward pressure, as covered in order flow.
Lead-lag between instruments. The most liquid instrument moves first. Index futures lead single stocks; the most active option leads the surface.
Momentum and reversion at different scales. Very short horizons often mean-revert (the bid-ask bounce and temporary liquidity pressure), while slightly longer ones can trend. Which regime applies is empirical and unstable.
Volatility signals. Volatility clusters, so recent volatility predicts near-term volatility. This does not predict direction but it does size spreads, which is where the money is.
What makes a signal real
Most apparent signals are artefacts of searching. Guarding against that requires a specific discipline, covered properly in p-values and multiple testing.
A mechanism. Why should this predict anything? Order book imbalance works because it reflects genuine supply and demand. A signal with no story is usually a coincidence with a good backtest.
Survival out of sample. On data never used in development, touched once.
Economic significance after costs. A signal predicting a move smaller than the spread is unprofitable no matter how statistically clean. See statistical vs practical significance.
Stability across regimes. A signal that only worked in one period probably fitted that period.
Signals in the wild are weak. A short-horizon predictor with 52% accuracy is genuinely valuable; anything claiming 70% is overfitted, mis-measured, or trading something that cannot be traded at scale.
Decay
Signals weaken over time as competitors find them and trade them away. This is not a possibility but a certainty, and it shapes how firms operate.
Consequences: research is continuous rather than a project, signals are monitored for degradation in live trading, and firms keep a pipeline rather than relying on any single one. A desk whose P&L depends on one signal is a desk with a countdown running.
Combining signals with quoting
Rather than trading a signal directly, most firms fold it into market making. The signal adjusts the four quoting controls:
Skew toward the predicted direction, so you accumulate inventory in the direction you expect to profit from.
Size larger on the side the signal favours.
Width tighter when signal confidence is high, wider when it is weak.
Participation: quote more actively when signals are strong, less when unclear.
This hybrid is strictly better than either pure approach. You keep earning the spread, which is a reliable income, while tilting the inventory you accumulate toward the direction with positive expectation.
"How would you use a signal that predicts the next price move?" A strong answer is not "buy when it says up". It is: skew quotes toward the prediction, so you get paid the spread as well as the signal, and you are not paying to cross when you could be earning.