Quantitative shops generally operate in one of two regimes. Regime A: a few strong alphas - a handful (3-10) of signals, each with a meaningful individual IC. The combination problem is conceptually simple - residualise, weight by significance, monitor each alpha individually. The risk is overfitting the weights to historical correlations. Regime B: many weak alphas - a large number (50-500+) of signals, each with a tiny individual IC. That combination problem becomes machine-learning territory (regularized regression, ensembles) and is beyond this chapter's scope; here we focus on Regime A, the typical setting for a working systematic trader.
Figure 1. Concept map for this episode: fight, desk translation, what to lock.
Weight stability over time
A subtle but critical issue: alpha weights should not change dramatically from one re-fit to the next. A pipeline that produces weights of roughly 50/30/20 this month and 10/70/20 next month is suffering from overfitting in the weight estimation - the weights are responding to noise in the recent data rather than to genuine changes in feature performance. The fix is shrinkage toward equal weights: blend the significance-implied weights with plain equal weights, using a shrinkage parameter that controls how much you trust the data versus the safe default. This is the alpha-weighting equivalent of the Bayesian shrinkage from Chapter 7 - equal weighting is the prior, IC-based weighting is the likelihood, and the blend is the posterior.
The trading-frequency choice
A subtle decision in alpha combination is how often to rebalance the composite alpha estimate itself. Re-estimating daily stays current with regime shifts, but the resulting weights are noisier and the composite alpha ends up more volatile than the underlying signals. Re-estimating quarterly, freezing weights in between, gives more stable weights and simpler production - but is slower to adapt to regime changes and can lock in bad weights for a whole quarter. The right answer depends on the strategy's horizon: multi-week strategies on relatively stable features typically re-estimate monthly or quarterly; sub-week strategies on fast-moving microstructure features re-estimate daily, with heavier shrinkage to compensate for the extra noise.
The bias-variance trade-off in alpha selection
A related strategic question: how many alphas to include in the combination. The naive answer is "all of them" - more features means more breadth, which improves FLAM. The realistic answer is "as many as you can validate." Each additional feature pays a multiple-testing tax (Chapter 15) and contributes a noise term that may dominate its signal. A practical rule of thumb: include only features whose residualised IC is significant after a Bonferroni correction sized to your full panel. Features that fail this test should be deprecated until the data supports them. Adding marginal features dilutes the composite; deprecating noise features sharpens it.
Worked picture
Residualised IC-implied weights: roughly 55%, 30%, and 15% across three alphas. Equal weights would give each about a third. Blending the two halfway gives a shrunk set of roughly 44%, 32%, and 24%. The dominant alpha keeps plurality but no longer receives 55% - an 11-point haircut that would have prevented a noisy IC spike from dominating the composite after one bad month.
What breaks in production
A common pattern in long-running quantitative shops: a feature that was added five years ago has been in the panel ever since, even though its rolling IC has trended toward zero. No one wants to be the one to deprecate it, because "we did the work to build it." The sunk cost is irrelevant. If the feature is no longer adding incremental IC, it is subtracting via the multiple-testing tax and the diluted composite. A disciplined panel review every six months should deprecate any feature whose 12-month residualised IC is not Bonferroni-significant. The discipline of regular pruning is what keeps the FLAM-implied IR realistic over multi-year horizons.
Key idea
Alpha combination is not about adding more signals; it is about combining the few good signals you have in a way that preserves their incremental information. The eleven-step engine, the shrinkage toward equal weights, and the discipline of pruning weak features are the three pillars of a durable combination process. Mastery of these three is worth more than any clever new signal.