Turn Databank candles into a first predictive model with scikit learn. Set a next bar target, build lagged return features that only see the past, split the series in time order with shuffle set to False so the test set stays in the future, and read the accuracy against an always one class baseline after costs. Validate the model walk forward with TimeSeriesSplit so it is scored across many periods and never trained on data that comes after its test fold.
