Signal Discovery

A trading signal is any piece of information that helps predict future returns, volatility or other outcomes: a price pattern, a valuation ratio, an analyst revision, a change in order flow or a shift in sentiment. Signal discovery is the process of generating ideas, turning them into measurable signals and testing whether they predict anything useful. Good signal research combines creativity with strict statistical discipline, because most candidate signals are noise.

Where signal ideas come from#

Signals with a clear reason to work are much more likely to survive out of sample.

Turning an idea into a signal#

1. Define precisely: for example, "change in consensus EPS estimate over the past month divided by price".
2. Ensure point in time availability. See Point-in-Time and Survivorship-Free Data.
3. Standardise: rank or z score across assets on each date. See Percentiles, Quantiles and Z-Scores.
4. Handle outliers: winsorise or rank. See Outliers and Robust Statistics.

Measuring predictive power#

Information coefficient (IC)#

The IC is the correlation between signal values and subsequent returns across assets on each date, often using rank correlation.

IC_t = rank correlation(signal_t, return_(t+1))

Even small ICs can be valuable when applied across many assets. Richard Grinold's fundamental law of active management links skill and breadth:

information ratio ≈ IC × √(breadth)

where breadth is the number of independent bets per year.

Quantile analysis#

Decay and horizon#

Signals predict best over certain horizons. Plotting IC against holding period shows how quickly information is absorbed: short term signals may decay within days, while value signals persist for months. Matching rebalancing frequency to signal decay keeps costs under control. See Signal Turnover, Breadth and Neutralization and Signal and Alpha Decay.

Avoiding false discoveries#

• Record every signal tested. See P-Hacking and Multiple Testing.
• Use out of sample periods. See In-Sample vs Out-of-Sample Testing.
• Control for known factors: a new signal may just repackage size, value or momentum. See Factor Models.
• Check robustness across universes and definitions.
• Watch for leakage. See Data Leakage.

Frequently asked questions#

What is a trading signal?#

A measurable piece of information, such as a valuation ratio, price pattern or estimate revision, that helps predict future returns or risk.

What is an information coefficient?#

The correlation between a signal's values and subsequent returns across assets, a common measure of a signal's predictive power.

How do I know if a new signal is real?#

It should have a sensible rationale, a consistent quantile pattern, positive out of sample results, survive costs and add value beyond known factors.

Next, learn why signals lose power over time in Signal and Alpha Decay.