Signal and alpha construction is the quantitative craft of turning raw data — prices, fundamentals, estimates, text, alternative data — into a numeric forecast of future cross-sectional returns, then refining and combining many such forecasts into a single tradeable score per asset. In the trade, a raw predictor is a signal; once it is cleaned, standardized, and scaled into expected-return units it becomes an alpha. The central tension of the discipline is that any signal strong enough to be tradeable is also weak enough to be mostly noise (even signals practitioners call "strong" typically have cross-sectional predictive correlations only around 0.05), and the search process that discovers signals is the same process that manufactures false ones — so most of the rigor lives not in finding alphas but in not fooling yourself about the ones you find.

How a signal becomes an alpha

The standard cross-sectional pipeline transforms a raw value (e.g. earnings yield, 12-month momentum) into a portfolio-ready score, step by step:

1. Cleaning / winsorization. Truncate or clip extreme cross-sectional outliers (commonly at the 1st/99th percentile or ±3 standard deviations) so a few mis-recorded values don't dominate. (FactSet) 2. Standardization. Convert each day's cross-section to a comparable scale via z-scores (subtract the cross-sectional mean, divide by the cross-sectional standard deviation) or percentile ranks. Ranking is more robust to non-normal, fat-tailed signals; z-scoring preserves magnitude information. 3. Neutralization. Regress the standardized signal on unwanted exposures — sector/industry dummies, size, beta — and keep the residual, so the bet is "company-specific" rather than an accidental sector or size tilt. (FactSet) 4. Scaling to alpha units. Grinold & Kahn's rule of thumb converts a standardized score into a residual-return forecast: alpha = volatility × IC × score, where IC is the information coefficient and volatility is residual return volatility. (Grinold & Kahn, Active Portfolio Management) 5. Combination. Blend multiple alphas (equal-weight, IC-weighted, or via a covariance-aware optimizer that accounts for correlation between signals) into one composite score that ranks the universe.

How it's used in practice

The yardstick for a single signal is the information coefficient (IC) — the cross-sectional correlation between the forecast and the subsequent realized return, measured period by period. IC near 0 means no skill; in equities a mean IC of roughly 0.03–0.06, sustained, is considered a genuinely good signal (a number quoted across practitioner texts, not a guarantee). Stability matters as much as level: the information ratio of the IC series itself (mean IC / standard deviation of IC, sometimes called the "IC IR") tells you whether the edge is dependable or lucky.

The organizing theory is Grinold's Fundamental Law of Active Management: the achievable information ratio is approximately IR ≈ IC × √BR, where breadth (BR) is the number of independent bets per year. (CFI, Grinold & Kahn) The law explains why quants prefer many small, weakly-predicted bets over a few high-conviction ones: a tiny IC, applied across thousands of independent positions, compounds into a respectable IR. The refinement adds a transfer coefficient (0–1) that discounts for the gap between the ideal portfolio and the one you can actually hold after constraints and costs: IR ≈ TC × IC × √BR. This is why signal construction can never be divorced from implementation — turnover, liquidity, and transaction costs erode raw IC into far smaller net alpha.

Adoption, debate & evidence

Signal/alpha construction is the foundational activity of the entire systematic-equity industry — quant funds, smart-beta products, and factor models all rest on it, and the IC / Fundamental-Law vocabulary is standard CFA and FRM curriculum. What is genuinely contested is how much of the published signal universe is real.

• The factor zoo. Harvey, Liu & Zhu (2016) catalogued 316 factors published in top journals (1967–2014) and argued that with so many tests, the conventional t > 2.0 hurdle is far too lax; they propose a multiple-testing threshold of roughly t > 3.0, under which a large share of published factors fail. (Harvey, Liu & Zhu, RFS 2016)
• Out-of-sample decay is real and large. McLean & Pontiff (2016) studied 97 published predictors and found returns ~26% lower out-of-sample and ~58% lower post-publication — evidence that some "alpha" was statistical mining and some was real but arbitraged away once known. (McLean & Pontiff, Journal of Finance 2016)
• Multi-signal overfitting compounds. Selecting the best k of n candidate signals inflates the apparent t-statistic so severely that valid critical values can be several times the textbook level; "highly significant" backtests are trivially manufacturable from pure-noise signals. (Alpha Architect)
• Alpha decay — the steady erosion of a signal's predictive power as it is discovered, crowded, and traded — is now treated as a structural feature, driving constant re-mining and the recent wave of ML/LLM alpha-mining systems. (AlphaAgent, arXiv 2025)

The honest summary: the machinery (IC, neutralization, the Fundamental Law) is sound and widely validated; the inventory of specific alphas is heavily populated by false or decayed discoveries.

Strengths & limitations

The framework's strength is that it makes "skill" measurable and decomposable: IC isolates forecasting quality, breadth isolates opportunity, transfer coefficient isolates implementation drag — so you can diagnose where a strategy fails. It works best in large, liquid, breadth-rich universes (cross-sectional equities) where the Law's diminishing √-breadth returns still pay off.

The limitations are sharp. The Fundamental Law assumes bets are independent — they rarely are; correlated signals overstate breadth and overstate the achievable IR, and independence "cannot be precisely measured without estimation error." (CFI) IC measured in-sample is upward-biased by the search that found it. The single biggest misuse is treating a backtested IC or Sharpe as the expected forward edge: between multiple-testing inflation, post-publication decay, and ignored costs, realized net alpha is routinely a fraction of the backtest — and sometimes zero or negative.

Sources

• Grinold, R. & Kahn, R., Active Portfolio Management (2nd ed.) — IC, alpha = vol·IC·score, Fundamental Law. Reference
• Corporate Finance Institute — Fundamental Law of Active Management, IR = IC·√BR, transfer coefficient, weaknesses. Link
• Harvey, Liu & Zhu (2016), "…and the Cross-Section of Expected Returns," Review of Financial Studies — 316-factor zoo, t > 3.0 threshold. Link
• McLean & Pontiff (2016), "Does Academic Research Destroy Stock Return Predictability?" Journal of Finance — 26% out-of-sample / 58% post-publication decay. Link
• FactSet, "A Granular Approach to Alpha Signal Selection and Optimization" — winsorization, standardization, neutralization workflow. Link
• Alpha Architect, "Backtesting Strategies Based on Multiple Signals — Beware of Overfitting Bias." Link

Dispute flags: the "good IC ≈ 0.03–0.06" range is a widely-cited practitioner heuristic, not a hard statistic. The exact survival count under t > 3.0 varies by sample/method across the Harvey-Liu-Zhu literature; decay magnitudes (26%/58%) are McLean-Pontiff point estimates with confidence intervals, not universal constants.