Machine Learning
Declarative DAG of ML components — features, targets, preprocessing, validation, models — trained, validated and materialized as a live indicator.
Premium module
The ML fluxo visual is a declarative DAG of components. Trains, validates and materializes the prediction as a derived series — the model becomes a live indicator, NO re-training at serving.
Backends
Pipeline components
feature_set selects columns. gerar_lags creates temporal lags. features_calendario adds hour/day (sin/cos). interagir_colunas combines features (ratio, product, difference).
target_direcao classifies +1/0/-1 with dead band. target_retorno continuous regression. target_conjunto multi-label categorical. target_custom labeling via Python script.
imputar (ffill, bfill, mean, median, constant). Scalers: zscore, minmax, robust, maxabs. winsorize clips outliers by quantile. reduzir_pca reduces dimensionality. selecionar_correlacao / selecionar_variancia filter features.
holdout simple temporal split. purged_kfold K-fold with embargo (prevents leakage). walk_forward expanding or rolling. custom partitioning via Python script.
modelo trains with a backend. modelo_custom trains via Python script. otimizar_hiperparametros grid or random search via cross-validation.
prever materializes prediction at ct://derived/name_pred. aplicar_modelo applies a trained model to a series without re-training — the model as a live indicator. avaliar computes custom metrics via Python script.
Economic evaluation
The fluxo visual can optionally run a teste de sobrevivência over the materialized prediction. Default: long if pred is positive, short if pred is negative. A custom custom strategy can be provided.