Alpha Research Process Explained

At a glance

Core measure: Information coefficient: signal versus subsequent return
Typical useful IC: 0.02 to 0.05
Key diagnostics: Decay profile, quantile monotonicity, turnover
Before strategy: Signal evaluation comes first, portfolio construction second

Key takeaways

Evaluate the signal before building the strategy: a feature that does not predict returns will not be rescued by clever position management.
The information coefficient measures the correlation between your signal and subsequent returns, and useful values are small.
The decay profile tells you the appropriate holding period: how quickly predictive power fades determines rebalancing frequency.
Quantile analysis reveals whether the relationship is monotonic or driven entirely by one extreme group.
Neutralising against known factors shows whether your signal contains anything beyond existing, freely available exposures.

The signal evaluation workflow

1
Construct the raw signal
A number for each instrument at each date, derived from the hypothesis. Handle missing data explicitly rather than filling silently.
2
Normalise cross-sectionally
Rank or z-score within each date so instruments are comparable and the signal is not dominated by scale differences.
3
Measure the information coefficient
The correlation between the signal and subsequent returns, computed per date and then averaged. Report the mean, the standard error, and the proportion of periods positive.
4
Plot the decay profile
IC at 1, 5, 10, 20, and 60 days ahead. This determines the natural holding period and whether the effect is immediate or gradual.
5
Run quantile analysis
Sort into five or ten buckets and compare average forward returns. A monotonic pattern is strong evidence; a result driven by one extreme bucket is fragile.
6
Neutralise against known factors
Regress out market, sector, size, value, and momentum. If the signal disappears, it is a repackaging of existing exposures.
7
Measure turnover
How much the signal changes between periods determines trading costs, which frequently decides viability.
8
Only then build a portfolio
Position sizing, constraints, and cost modelling come after the signal has demonstrated predictive content.

The information coefficient in practice

For each date t:
   IC(t) = Spearman correlation between
           signal values across instruments at t
           and forward returns from t to t+h

Report:
   mean IC          typical useful range 0.02 to 0.05
   IC standard dev  measures stability
   IC information ratio = mean IC / std dev of IC
                    values above 0.3 are respectable
   % periods positive   above 55% suggests consistency

Why such small numbers are useful:
   Expected IR ~ IC x sqrt(breadth)
   IC of 0.03 across 300 names rebalanced monthly:
      breadth ~ 3,600 -> IR ~ 1.8

A single-name forecast with IC 0.03 is worthless.
The same signal across hundreds of names is a business.

Computing and interpreting the IC.

Decay analysis

How quickly a signal loses predictive power determines almost everything downstream: rebalancing frequency, turnover, cost sensitivity, and capacity.

Horizon	Fast decay	Medium decay	Slow decay
1 day	0.045	0.020	0.008
5 days	0.018	0.022	0.010
20 days	0.004	0.019	0.014
60 days	0.001	0.009	0.015
Implication	Trade daily; costs dominate	Rebalance monthly	Quarterly; low turnover
Capacity	Low	Moderate	High

Three decay profiles and what each implies.

Fast-decaying signals are usually microstructure effects that require excellent execution to capture. Slow-decaying signals tolerate higher costs and larger capital but require patience through long periods where the signal appears not to work.

Neutralisation: is the signal new?

Many apparently novel signals are proxies for well-known factors. A signal correlated with size, value, or momentum is providing exposure available in low-cost funds, which is useful to know before building a strategy around it.

Step 1: raw signal IC = 0.041

Step 2: regress the signal on known factors
        signal ~ market_beta + size + value + momentum + sector
        take the residual

Step 3: recompute IC using the residual signal
        residual IC = 0.009

Interpretation:
   Roughly 80% of the predictive power came from
   exposure to known factors.  The signal is mostly
   a repackaged momentum tilt.

   That is not useless information: it tells you what
   you are actually buying, and that the same exposure
   is available more cheaply elsewhere.

If residual IC remains close to the raw IC, the signal
contains something the known factors do not capture,
which is the interesting case.

Testing whether a signal survives factor neutralisation.

Combining signals

Check correlation between signals first. Two signals correlated at 0.8 are one signal with extra computation and extra fitting risk.
Start with equal weighting. It is remarkably difficult to beat, and it requires estimating nothing, which removes a major source of overfitting.
Weight by IC and correlation only with a long history. Optimised weights fitted on a short sample are noise dressed as sophistication.
Consider signal-conditional weighting. Some signals work better in specific regimes; conditioning on volatility or trend state is a middle path.
Watch aggregate turnover. Combining a slow and a fast signal can produce turnover close to the fast one, inheriting its cost profile.
Validate the combination separately. A combination of validated signals is itself a new model that has been fitted, and requires its own out-of-sample test.

Frequently asked questions

What is a good information coefficient?

Between 0.02 and 0.05 for a cross-sectional equity signal is genuinely useful when applied across many instruments. Values above 0.10 on a liquid universe warrant suspicion rather than celebration, since they more commonly indicate look-ahead bias or a data error than a discovery.

Why evaluate the signal before building a strategy?

Because a feature with no predictive content cannot be rescued by position sizing or exit rules, and building the full strategy first makes it impossible to tell whether a poor result came from the signal or the implementation. Signal evaluation is fast, cheap, and eliminates most ideas before expensive work begins.

What does signal decay tell me?

The appropriate holding period and rebalancing frequency. A signal with predictive power that vanishes within two days requires daily rebalancing and excellent execution; one that persists for months tolerates quarterly rebalancing and much higher costs. Decay determines capacity as much as it determines frequency.

Should I neutralise my signal against factors?

At minimum, test whether it survives neutralisation. If most of the predictive power disappears, the signal is providing known factor exposure, which is available more cheaply elsewhere. Whether to neutralise in the live strategy depends on whether you want that factor exposure deliberately.

How many signals should a strategy combine?

Three to ten genuinely uncorrelated signals is a practical range. Beyond that, additional signals are usually correlated with existing ones, and the weighting problem becomes a fitting exercise. Breadth across instruments contributes far more than breadth across signals.

Test this idea before you trade it

Describe the rules in plain language and AlgoTrader AI turns them into a structured strategy blueprint with a configurable historical backtest, cost assumptions, and exportable code.

Build a backtest

Keep reading

Referenced by

Educational use only. This guide explains how a strategy works. It is not investment advice, not a recommendation, and no result described here is a forecast. Test any approach on historical and out-of-sample data, size positions conservatively, and never risk money you cannot afford to lose.