Signal Construction

Signal construction is the craft that sits between a research idea and a P&L. A raw predictor, analyst revisions, a valuation ratio, an order-flow imbalance, is never tradeable as-is: it has outliers, unwanted exposures, wrong scale, and unknown quality. Turning it into a clean, comparable, risk-controlled signal, and measuring how good that signal is, is the daily work of a quant researcher, and the fundamental law of active management is the theory that governs it.

From raw predictor to signal

The pipeline is a sequence of transformations, each with a purpose:

1. Clean and winsorize. Cap extreme values (e.g., at the 1st/99th percentiles) so a single data error or outlier does not dominate. Squared-loss objectives and covariance estimates are exquisitely sensitive to outliers (Ordinary Least Squares (OLS)).
2. Standardize cross-sectionally. Convert the raw predictor $c_i$ into a unit-free score, either a z-score or a rank:

$z_i = \frac{c_i - \operatorname{median}(c)}{\operatorname{MAD}(c)} \quad\text{or}\quad z_i = \Phi^{-1}\!\Big(\frac{\text{rank}(c_i)}{N+1}\Big).$

Ranking is robust to outliers and monotone transforms but throws away magnitude; z-scoring keeps magnitude but assumes the cross-section is roughly symmetric. Most desks rank-transform, then map to a normal score. 3. Neutralize unwanted exposures. Regress the signal on things you do not want to bet on, market beta, sector dummies, size, and keep the residual:

$\tilde z_i = z_i - \hat z_i, \qquad z = X\hat\gamma + \tilde z.$

The residual $\tilde z$ is orthogonal to $X$ by the Ordinary Least Squares (OLS) projection property, so the signal no longer secretly loads on sector or size, it is pure alpha in the intended dimension. This is how you avoid a "value" signal that is really an energy-sector bet. 4. Combine and scale. Blend multiple signals (weighting by conviction or via a covariance-aware combination) and scale to a target risk or a target position.

Grinold's alpha: from score to expected return

A z-score is dimensionless; to size positions you need an expected return. Grinold's formula converts a standardized signal into an alpha forecast:

$\alpha_i = \text{IC} \times \sigma_i \times z_i,$

where $\sigma_i$ is the asset's residual volatility and IC is the information coefficient (below). The logic: your best guess of the residual return is your signal, shrunk by how predictive the signal actually is (IC) and scaled by how much the asset can move ($\sigma$). A signal that is only 5% correlated with future returns should drive small positions, however extreme its z-score, the shrinkage by IC is the discipline that prevents overbetting a weak edge.

The information coefficient

The information coefficient is the cross-sectional correlation between the signal at time $t$ and realized forward returns:

$\text{IC}_t = \operatorname{corr}\big(z_{i,t},\, r_{i,t+1}\big).$

It is the single most important diagnostic of signal quality. Realistic equity ICs are small, a monthly IC of 0.03–0.05 is a genuinely good signal, and 0.10 is exceptional. The average IC measures edge; the volatility of IC measures how reliable that edge is; the IC information ratio, $\overline{\text{IC}}/\sigma(\text{IC})$, is what actually matters. Because a single-period IC is tiny and noisy, it is only tradeable when spread across many independent bets, which is the fundamental law.

The fundamental law of active management

Grinold's law connects skill and breadth to the achievable Information Ratio:

$\boxed{\text{IR} = \text{IC} \cdot \sqrt{\text{BR}},}$

where BR is breadth, the number of independent bets per year. The derivation is just the $\sqrt{n}$ law of large numbers: each bet contributes a tiny IC of edge with independent noise, so averaging $N$ of them scales the signal-to-noise by $\sqrt{N}$. The law is the most important sentence in quant portfolio management because it says a small edge, applied widely and independently, is a great strategy. An IC of 0.05 over 500 stocks rebalanced monthly (BR ≈ 500 × 12 = 6000, if independent) gives $\text{IR} = 0.05\sqrt{6000} \approx 3.9$, spectacular. The same IC on 5 macro bets a year gives $\text{IR} = 0.05\sqrt{5} \approx 0.11$, untradeable. This is why stat-arb runs thousands of names and macro runs few.

Clarke, de Silva, and Thorley refined it with a transfer coefficient $\text{TC}$, the correlation between your ideal and your actual (constrained) portfolio: $\text{IR} = \text{TC}\cdot\text{IC}\cdot\sqrt{\text{BR}}$. Long-only constraints, position limits, and transaction costs all shrink TC below 1, so real IRs fall well short of the frictionless law.

Worked example: how many stocks do you need?

You have a signal with monthly IC = 0.03 and you rebalance monthly. To hit a target IR of 2.0, the law requires $\sqrt{\text{BR}} = \text{IR}/\text{IC} = 2/0.03 = 66.7$, so $\text{BR} = 4444$ bets/year, i.e. about $4444/12 \approx 370$ independent names per month. The word "independent" is load-bearing: if your 370 stocks are really 10 correlated sector clusters, effective breadth collapses and your realized IR is a fraction of the promise. Estimating effective breadth under correlation, not raw name count, is where the law meets reality.

Failure modes

• Overstated breadth. Correlated bets are not independent; effective breadth is far below the number of positions, so the law flatters undiversified books.
• Look-ahead and survivorship. Using data not available at signal time (unlagged fundamentals, restated financials) inflates IC in backtests and evaporates live.
• IC instability and decay. IC is regime-dependent and shrinks as signals crowd; a signal with mean IC 0.05 but huge IC volatility is barely tradeable (Alpha Decay).
• Neutralization removes the alpha. Over-neutralizing (stripping every exposure) can orthogonalize away the very thing that pays; you must neutralize risk, not return.
• Combination overfitting. Optimizing signal weights on the same sample that discovered them is Overfitting; use shrinkage or equal-weighting.

In interviews

Walk the pipeline: winsorize, rank/z-score cross-sectionally, neutralize unwanted exposures via regression residuals, then scale by IC and vol into an alpha. Define the information coefficient as the cross-sectional correlation of signal with forward returns and stress that realistic ICs are tiny (0.03–0.05). The centerpiece is the fundamental law $\text{IR} = \text{IC}\sqrt{\text{BR}}$, derive it as a $\sqrt{n}$ effect and use it to explain why a weak signal over thousands of independent names beats a strong signal over a handful of bets. The best candidates immediately flag that "breadth" means independent bets and that correlation destroys effective breadth. See Cross-Sectional vs. Time-Series Strategies for how breadth differs between the two strategy families.