Recency Weighting and Decay in Training Samples
Markets change regime over time, so a model that treats a sample from five years ago as equally informative as one from last week is implicitly assuming nothing important has changed — recency weighting lets recent data count for more.
A standard model trains by minimizing average error across all training samples, giving each one equal say regardless of when it occurred. But market relationships drift — the way a value signal behaved in a low-rate environment may not hold in a high-rate one, and the microstructure of a market five years ago may differ meaningfully from today's. Recency weighting assigns each training sample a weight that decays with age, typically exponentially, so a sample from last month pulls the model harder than an otherwise-identical sample from three years ago, while very old samples still contribute a little rather than being discarded outright.
The decay rate is itself a parameter to tune: too fast, and the model overreacts to short-lived noise and effectively trains on a tiny recent window; too slow, and it barely differs from equal weighting and stays slow to adapt when a regime genuinely shifts.
Exponential recency weighting lets a model adapt to regime change by favoring recent data, but the decay rate trades off responsiveness against the whiplash of overreacting to short-term noise.
With a half-life of 250 trading days, a sample from exactly one year ago carries half the weight of a sample from today, and a sample from two years ago carries a quarter — old data is never zeroed out, just steadily discounted.
Practice in interviews
Further reading
- de Prado, Advances in Financial Machine Learning (ch. 4)