MOEX Price Prediction

MOEX Price Prediction is an end-to-end forecasting solution for Moscow Exchange (MOEX) stocks (e.g., SBER, GAZP, ROSN) using historical data, technical indicators, and deep learning with PyTorch.

---

Table of Contents

1. Project Overview
2. Key Features
3. Architecture & Structure
4. Technologies
5. Hydra Configuration
6. Installation & Setup
7. Usage
8. API Reference
9. Model Training & Visualization
10. License

---

Project Overview

This project fetches MOEX historical end-of-day data (for any ticker available via the MOEX API in app/data.py), computes technical indicators (RSI, SMA, MACD, Bollinger Bands, ATR), trains PyTorch models (LSTM-Attention, TCN, Transformer) via Hydra-powered experiments, and serves predictions through a FastAPI REST API. See Hydra Configuration for details.

---

Key Features

• Modular Architecture: Clear separation between data ingestion, preprocessing, model training, and serving
• Hydra Experiments: Easily switch models (lstm, tcn, tft) and tickers (SBER, GAZP, ROSN, etc.) with command-line overrides
• Versioned Artifacts: Models and scalers saved under saved_models/v{version}; metadata tracks versions and architecture
• Auto-Retraining: Optional performance monitoring triggers retraining via MLflow and Optuna
• Live API: FastAPI endpoint for on-demand predictions

---

Architecture & Structure

MOEX_PREDICT/
├── app/
│   ├── data.py                 # MOEX & CBR data loader + DataLoader factory
│   ├── preprocessing.py        # Indicator calculations
│   ├── models/                 # Model definitions and factory
│   │   ├── attention_lstm.py
│   │   ├── tcn.py
│   │   ├── tft.py
│   │   └── factory.py
│   ├── model_manager.py        # Loading versioned models & scalers
│   ├── transfer_learning.py    # Retraining logic & metadata
│   ├── predict.py              # Prediction wrapper
│   ├── monitoring.py           # Performance validation
│   └── main.py                 # FastAPI application
├── conf/                       # Hydra configuration
│   ├── config.yaml
│   ├── data/
│   │   └── default.yaml        # Data loader settings
│   ├── model/
│   │   ├── lstm.yaml           # LSTM params
│   │   ├── tcn.yaml            # TCN params
│   │   └── tft.yaml            # Transformer params
│   ├── optimization/
│   │   └── default.yaml        # HPO settings
│   └── train/
│       └── default.yaml        # Training settings & versioning
├── saved_models/               # Versioned model artifacts
│   ├── v1/
│   │   ├── SBER_model.pth
│   │   ├── SBER_scaler_X.pkl
│   │   ├── SBER_scaler_y.pkl
│   │   └── ...
│   └── v2/
│       └── ...
├── train.py                    # Hydra entrypoint for experiments
├── Makefile                    # install, run, clean commands
└── README.md                   # This file

---

Technologies

• Python: 3.11
• Config: Hydra (1.3.2), OmegaConf
• Data: Pandas, NumPy, scikit-learn
• Deep Learning: PyTorch 2.6
• Experiment Tracking: MLflow 2.21
• Web API: FastAPI, Uvicorn
• Scheduling: APScheduler

---

Hydra Configuration

All experiments are driven by conf/config.yaml. Override sections via CLI:

# Train TCN on GAZP, version v2
python3 train.py \
    model=tcn \
    data.ticker=GAZP \
    train.horizon=5 \
    train.version=v2

# Train model with HPO
python3 train.py \
    model=lstm \
    data.ticker=SBER \
    data.start_date=2013-01-01 \
    train.horizon=5 \
    train.epochs=20 \
    train.version=v1 \
    optimization.enable=true \
    optimization.n_trials=20 \
    optimization.epochs_per_trial=20

See conf/ for defaults.

---

Installation & Setup

# 1. Clone repository
git clone https://github.com/NasdormML/Moex_predict.git
cd Moex_predict

# 2. Create virtual environment and install dependencies
make install

# 3. Run MLflow and FastAPI
make run

• MLflow UI: http://127.0.0.1:5001
• API Docs: http://127.0.0.1:8000/docs

Available Make Commands

Command	Description
make install	Create venv and install dependencies
make run	Run MLflow + FastAPI (development with reload)
make run-prod	Run MLflow + FastAPI (production with workers)
make api	Run only FastAPI
make mlflow	Run only MLflow server
make lint	Run Flake8 linter
make clean	Remove __pycache__ and *.pyc files

---

Usage

1. Train a Model

python train.py model=lstm data.ticker=SBER

2. Get Predictions via API

curl -X POST "http://127.0.0.1:8000/predict/SBER/2026-02-10" \
    -H "Content-Type: application/json"

---

API Reference

POST /predict/{ticker}/{target_date}

Generate price forecast for a ticker up to the target date.

Parameters:

Name	Type	Description
ticker	string	Stock ticker (e.g., SBER, GAZP)
target_date	date	Forecast horizon limit (YYYY-MM-DD)

Response:

{
  "ticker": "SBER",
  "known_up_to": "2026-02-03",
  "requested_target_date": "2026-02-10",
  "forecast_dates": [
    "2026-02-04",
    "2026-02-05",
    "2026-02-06",
    "2026-02-07",
    "2026-02-10"
  ],
  "predictions": [
    307.40,
    306.08,
    305.13,
    305.05,
    305.28
  ]
}

GET /health

Health check endpoint.

Response:

{
  "status": "healthy",
  "models_loaded": 1
}

---

Model Training & Visualization

SBER Performance

Model	MSE	RMSE	MAE	MAPE
LSTM-Attn	19.87	4.46	3.11	1.20%
Transformer	11.37	3.37	2.42	0.97%

GAZP Performance

Model	MSE	RMSE	MAE	MAPE
TCN	18.31	4.28	3.04	2.08%

ROSN Performance

Model	MSE	RMSE	MAE	MAPE
Transformer	61.50	7.84	5.65	1.10%

---

License

This project is licensed under the MIT License.