optuna/optuna

What it is

A hyperparameter optimisation framework. You write a function that scores one configuration; Optuna chooses which configurations to try, using the results so far to sample the promising regions, and prunes runs that are clearly going nowhere before they finish. Every trial is stored in a database, so a study can be paused, resumed and inspected.

What it is good for. The real constraint in strategy research is how many hypotheses you can kill per day, and grid search is the slowest possible way to kill them: entry threshold × stop × target × session, and a full sweep runs thousands of backtests, most of them on parameter values that were obviously bad after the first six months of data. Optuna's sampler concentrates on the regions that are working, and its pruning stops a hopeless trial part-way — together that is usually an order of magnitude fewer full backtests for the same answer, on a CPU box that has other jobs to do. It also fits the tools already in the ledger rather than replacing them: it wraps skforecast's rolling-origin backtest as the objective function, so the thing being optimised is walk-forward performance rather than a curve fitted to the whole history. One warning, and it is the important part: a smarter search finds an overfitted result faster than a dumb one. Everything Edition 6 said about arch's reality check applies double here — the more configurations you try, the more certain it is that the best of them is luck. Optuna finds candidates; it never validates them.