Prediction Models

Prediction Models is a model zoo for common ML tasks: ARIMA/Prophet/XGBoost for forecasting, XGBoost/LightGBM/CatBoost for tabular, CNN/Transformer for sequences. Unified FastAPI serving, ONNX export, and automated benchmarking against baselines.

Teams reimplement standard models repeatedly. This provides battle-tested, benchmarked models with consistent interfaces — drop in, configure, serve.

The repository evidence points to these observable boundaries and technologies:

• Python package (prediction_models/)
• Modular model adapters (sklearn, xgboost, pytorch)
• FastAPI serving with /predict, /batch endpoints
• ONNX export for cross-platform inference
• Pytest benchmarks with synthetic + real datasets
• Docker multi-stage for CPU/GPU images

• Unified BasePredictor interface (fit, predict, save, load)
• Forecasting: ARIMA, Prophet, XGBoostTS, NBEATS
• Tabular: XGBoost, LightGBM, CatBoost, TabPFN
• Sequence: 1D-CNN, Transformer, LSTM
• ONNX export for all models
• Benchmark suite: MASE, RMSE, MAE on standard datasets

• Dependency conflicts across model libraries
• ONNX export coverage for gradient boosting
• Benchmark reproducibility across environments
• Model versioning and registry integration

• All model adapters implement BasePredictor
• FastAPI serving handles batch + streaming
• ONNX export working for sklearn/XGBoost
• Benchmarks run in CI with cached datasets

Model zoos need strict interface contracts. ONNX is great for sklearn but limited for boosting. Consider MLflow for registry instead of custom versioning.