An investment signal is a quantitative measure derived from data that predicts future asset returns with better-than-random accuracy. A momentum signal might be the 12-1 month return. A quality signal might be gross profitability (gross profit divided by assets). A sentiment signal might be the aggregate analyst revision ratio (upgrades minus downgrades). Each signal has a characteristic: cross-sectional (ranks stocks relative to each other) or time-series (ranks a stock relative to its own history); slow-moving (fundamental signals update quarterly) or fast-moving (price-based signals update continuously).

Signal decay — how quickly predictive power fades after the signal is computed — determines the appropriate holding period and rebalancing frequency. Price-based signals (momentum) have predictive power at the monthly to quarterly frequency; they add little incremental predictive power after 12 months. Analyst revision signals are meaningful over weeks to months. Fundamental valuation signals have the longest decay — a stock that is cheap today may remain cheap for years, but the valuation premium eventually materializes. Matching rebalancing frequency to signal decay minimizes unnecessary turnover while maintaining full signal exposure.

Combining multiple signals typically improves predictive accuracy more than optimizing any single signal in isolation — a core insight from ensemble methods and multi-factor investing alike. The benefit of combination comes from uncorrelated errors: when Signal A makes an incorrect prediction about stock X, Signal B is likely to be wrong about a different stock Y, so their combined prediction is more accurate than either alone. The correlation between signals determines the combination benefit — the lower the pairwise correlation between signals, the greater the diversification benefit of combining them.

Ensemble approaches include simple averaging (equal-weight each signal), rank-based combination (average the percentile ranks from each signal), or supervised ML weighting (using historical data to learn the optimal weights). Research consistently shows that simple equal-weight combination of well-chosen signals is surprisingly competitive with sophisticated ML-weighted combinations — consistent with the overfitting risk of estimating weights from historical data. The most important input to signal combination is signal selection (choosing signals that are genuinely predictive and genuinely independent) rather than the weighting methodology.

Signal predictive power varies across market regimes. Value signals work better in early-cycle recoveries than in late-cycle growth environments. Momentum signals are stronger in low-volatility trending markets than in high-volatility reversal-prone environments. Quality signals provide more alpha during recessions (when financial distress risk is elevated) than in expansions (when even weak-balance-sheet companies survive). Regime-aware signal filtering adjusts signal weights based on the current macroeconomic or market environment to improve overall composite signal accuracy.

Practical regime filters: VIX level (high VIX → reduce momentum weight, increase quality weight), yield curve slope (inverted curve → reduce value weight, increase defensive signals), economic expansion/contraction signals (PMI, leading indicators), and volatility regime detection (realized volatility relative to 12-month average). These filters add complexity and data-fitting risk — each regime classification is itself a model with its own estimation error. The benefit is most clear-cut for extreme regimes; nuanced regime classifications often overfit and add noise rather than signal.