Model Training
LTprophecy provides a managed ML training pipeline supporting XGBoost, LightGBM, CatBoost, Prophet, and ensemble methods with automated hyperparameter optimization via Optuna.
Available Algorithms
| Algorithm | Best For | Plan |
|---|---|---|
| Prophet | Seasonal time series with holiday effects | All |
| XGBoost | Tabular data with many features | All |
| LightGBM | Large datasets, faster training | Growth+ |
| CatBoost | High-cardinality categoricals | Growth+ |
| Ensemble | Maximum accuracy via model stacking | Enterprise |
| Custom (BYOM) | Your own scikit-learn pipeline | Enterprise |
Training Configuration
Basic Options
- Algorithm — select from the table above
- Dataset & version — pin to a specific dataset version
- Target column — the numeric column to forecast
- Feature columns — additional inputs to the model
- Validation split — train/val/test ratio (default 70/15/15)
- Forecast horizon — number of periods to forecast ahead
Hyperparameter Optimization
Enable Auto-tune (Optuna) to let the platform search for optimal hyperparameters. Configure:
- n_trials — number of Optuna trials (default: 50)
- timeout_minutes — maximum search time
- Metric — RMSE, MAE, MAPE, or sMAPE
- CV folds — time-series cross-validation folds (default: 5)
- Pruning — early stopping of unpromising trials
Training Jobs & GPU Queue
Training jobs are queued through Celery workers. CPU workers handle most models; GPU workers (if provisioned) accelerate deep learning and large LightGBM runs. You can monitor job status in real-time via Models → Training Runs.
Typical training times:
- Prophet (≤ 100k rows): < 2 min
- XGBoost with Optuna (100k rows, 50 trials): 5–15 min
- Ensemble stack: 20–45 min
Experiment Tracking (MLflow)
Every training run is automatically logged to MLflow with:
- All hyperparameters
- Evaluation metrics (RMSE, MAE, R², etc.)
- Feature importance plots
- Confusion matrices and residual charts
- Serialized model artifact (stored in MinIO)
Access the MLflow UI via Models → MLflow Dashboard (admins only in production).
Model Registry & Promotion
After training, models pass through lifecycle stages:
- Staging — trained, under evaluation
- Production — promoted by an admin, used for forecasts
- Archived — retired, artifacts retained
Only Production models can be selected when creating new forecasts. Promoting a model requires org:models:manage permission.
Model Evaluation
The evaluation panel shows:
- Hold-out test set metrics
- Backtested forecasts vs actuals chart
- Shapley value feature importance (XGBoost/LightGBM)
- Drift detection against training data distribution
- Calibration curves for probabilistic forecasters