Ensembling Across Seeds and Refit Dates
A single trained model is one noisy draw from a process with randomness in initialization and in the exact data cutoff used, and averaging predictions across many such draws produces a steadier, better-calibrated signal.
Train the same gradient-boosted model twice with different random seeds, and you'll get two slightly different sets of predictions, because random seeds control things like data shuffling, feature subsampling, and initial splits. Train it on data ending last Friday versus data ending the Friday before, and you get another source of variation, since each refit sees a slightly different history. Neither version is "the" model — each is one noisy sample from a distribution of plausible models the same pipeline could have produced. Ensembling across seeds and refit dates means training several versions that differ only in these incidental choices and averaging their output, which cancels out the noise specific to any one run while keeping the signal common to all of them.
This is distinct from ensembling different model types (say, boosting plus a neural network): seed-and-refit ensembling holds the modeling approach fixed and only varies the randomness and the training cutoff, so it targets variance reduction specifically rather than trying to combine different sources of skill.
Averaging predictions across multiple seeds and refit dates of the same model removes noise that is specific to any single training run, leaving a steadier estimate of what the model actually believes.
If five seeds of the same model predict next-day returns of 0.8%, 1.3%, 0.6%, 1.1%, and 0.9% for a stock, the ensemble average of about 0.94% is a more stable estimate than trusting any single seed's number.
Practice in interviews
Further reading
- de Prado, Advances in Financial Machine Learning (ch. 8)