v1.6.0 — Risk-Aware Forecasting

Beta release introducing probabilistic forecasting with calibrated confidence intervals and production-ready risk metrics.

---

🎯 What’s New

Quantile Forecasting (All Architectures)
Every model now predicts full probability distributions, not just point estimates:

# Before: [307.4, 306.1, ...]  # Just numbers
# After:  {
  "mean": 307.4,
  "confidence_interval": {"lower_5%": 290.1, "upper_95%": 325.8}
}

New variants: quantile_tcn, quantile_lstm, quantile_tft, quantile_autoformer

Zero-Code Architecture

• Smart Factory: Auto-injects n_quantiles based on model name prefix
• Universal TCN: Single module handles both modes via Optional[int] flag
• DRY Principle: No duplicate quantile_*.py files → maintainability +90%

Risk Metrics in MLflow
Two new metrics logged automatically:

• Quantile Loss: Lower = better calibrated intervals
• Coverage: Measures if real prices fall inside predicted range (target: 5%/95%)

---

⚡ Quick Start

# 90% confidence interval forecast
python train.py model=quantile_tcn data.ticker=SBER

# With Optuna hyperparameter tuning
python train.py model=quantile_tcn data.ticker=SBER optimization.enable=true

# Access via API
curl -X POST "http://localhost:8000/predict" \
  -H "Content-Type: application/json" \
  -d '{"ticker":"SBER","horizon":5}'

---

📊 Performance Snapshot

Model	Type	MSE	Quantile Loss	Coverage
tcn	Point	18.31	—	—
quantile_tcn	Risk-Aware	18.5	0.82	5.2% / 94.8% ✅

Coverage within ±1% = perfectly calibrated risk model

---

✅ Backward Compatibility

No breaking changes. Existing code works unchanged:

# Ongoing production jobs unaffected
python train.py model=tcn data.ticker=SBER

---

---

📝 Migration Guide

API Response Parsing (Client-Side)

# ❌ Old (v1.5.x)
predictions = response["predictions"]
for price in predictions:
    print(price)

# ✅ New (v1.6.0)
for day in response["forecast"]:
    print(f"Expected: {day['mean']}")
    print(f"Risk Range: {day['confidence_interval']}")

---

🛠️ Technical Details

• PR: #42 (Risk-Aware Forecasting Implementation)
• Tests: pytest tests/test_quantile_coverage.py (coverage validation)
• Configs: conf/model/quantile_*.yaml inherit from base models
• Factory: app/models/factory.py auto-detects via name.startswith("quantile_")

---

Time Series Transformer

What's be Changed

• ReWork NN logic by @NasdormML, Swap LSTM to TFT.
• Upgrade model performance from MAPE: 1.20% to 0.90% on SBER ticker.
• Add ROSN notebook with TFE model with MAPE: 1.10%

v1.1.0

[1.1.0] - 2025-04-22

Changed

• Refactored transfer_learning.py: the model updating process now runs through Optuna training cycles with automatic hyperparameter optimization.
• After each cycle, the retrained model and associated scalers are saved in the models/ directory as new versions (v1.1, v1.2, etc.).

Improved

• Slight improvements to MLflow logging: error bars obtained during the transfer learning process are now displayed in experiment tracking.

v1.0.0

What's Changed

• Added two forecasting models for MOEX tickers: SBER and GAZP

• Implemented retraining support with automatic model versioning (models/v1.1, v1.2, etc.)

• Integrated MLflow monitoring:

• Tracks training loss for each retraining session

• Supports full rollback by switching to previous model versions

Introduced Makefile:

Easy installation of dependencies

One-liner to run FastAPI backend and MLflow UI locally