Machine learning for trading
without writing Python.
Visual declarative pipeline: drag, connect, configure. Feature engineering, temporal cross-validation, hyperparameter tuning and deployment — all in the same app, all composable.
Verida ML — visual declarative DAG, from raw data to materialized prediction
The pipeline flow
From raw data to materialized prediction as a live indicator.
Model backends
Built-in or custom. The framework is algorithm-agnostic.
Gradient Boosting
gbdt Decision trees in ensemble. Robust, handles non-linearity. Via scikit-learn.
MLP
mlp Multilayer neural network. Via PyTorch. Flexible for complex patterns.
Random Forest
random_forest Ensemble of independent trees. Reduces variance, resistant to overfitting.
Centroid
centroide Centroid distance classifier. Simple, interpretable, no external dependencies.
KNN
knn K-Nearest Neighbors. Classification by proximity in feature space.
K-Means
kmeans Unsupervised clustering. Groups patterns without labels, useful for regimes.
Isolation Forest
isolation_forest Anomaly detection. Identifies outliers without prior labeling.
Linear
linear Linear regression/classification. Interpretable baseline, fast to train.
+ Custom backends via inline Python script — define treinar/inferir and use any library.
DAG components
Each node is a transformation. Compose as you wish.
Feature Engineering
Lags, transforms (log, sqrt, return_pct), interactions (ratio, product, difference).
Calendar Features
Hour of day, day of week, sinusoidal encoding for cyclicality.
Imputation
NaN filling (mean, median, zero, ffill, bfill). Causal, no look-ahead.
Outlier Removal
Clipping by empirical quantiles. Removes extremes without distorting distribution.
Variance Threshold
Discards features with zero variance. Eliminates columns that add no information.
Correlation Filter
Keeps top-k features by correlation with target. Reduces multicollinearity.
Cross-validation
Temporal k-fold, walk-forward (expanding or rolling), with embargo.
Hyperparameter Tuning
Grid search or random search, with ranking metric evaluation.
Labeling
Labels as direction (+1/0/-1), continuous return, or categorical combination.
ML is not an oracle
The pipeline optimizes on historical data. Better than intuition? Maybe. Better than manual overfitting? Certainly. Predicts the future? No. Markets are unknowable.
What ML does: finds patterns in the past. What you do: decide if those patterns matter.