Time Series Forecasting Lab

A practical and research-oriented guide to modern time-series forecasting, with a focused financial time-series seminar track.

This repository covers the full journey from classical statistical models such as Moving Average, ARIMA, and Prophet to recent deep learning and foundation models including DLinear, PatchTST, TimeMixer, iTransformer, CycleNet, TimesFM, SparseTSF, and TS-RAG.

The goal is not only to reproduce papers, but also to understand how to define forecasting problems, build reliable baselines, evaluate models correctly, and apply time-series forecasting techniques in real-world business, production, and financial-market environments.

Who This Is For

• MLOps Engineers who want to operationalize forecasting systems
• ML Engineers and Data Scientists who need reliable baselines
• AI Research Engineers who want a structured reading path for modern time-series models
• Backend/Data Engineers who need to understand forecasting workflows end to end
• Financial time-series students who need leakage-aware forecasting and backtesting discipline

Learning Path

1. Define the forecasting problem before choosing a model
2. Analyze time-series data and identify seasonality, trend, anomalies, and missing intervals
3. Build simple but strong baselines
4. Compare statistical, ML, DL, and foundation models
5. Avoid data leakage and invalid validation methods
6. Understand modern forecasting research through financial time-series constraints
7. Prepare models for production with retraining, monitoring, risk checks, and drift checks

Repository Structure

ts-forecasting-lab/
├── docs/                 # Conceptual guides and production checklist
├── notebooks/            # Experiment notebooks placeholders
├── src/ts_forecasting/   # Minimal Python package skeleton
├── configs/              # Model and experiment configs
├── experiments/          # Dataset-specific experiment areas
├── papers/               # Paper reading notes
├── seminar/              # Five-day financial time-series seminar
├── docker/               # Container environment
├── tests/                # Unit tests
├── Justfile
└── pyproject.toml

Model Map

Level	Category	Models
0	Naive Baselines	Last Value, Seasonal Naive, Moving Average
1	Statistical Models	ARIMA, SARIMA, Exponential Smoothing, Prophet
2	ML/DL Models	LightGBM with lag features, LSTM, DLinear, PatchTST
3	Modern Forecasting Architectures	TimeMixer, iTransformer, CycleNet, SparseTSF
4	Foundation / Retrieval Models	TimesFM, Chronos, Moirai, TS-RAG

Recommended Datasets

• Electricity: demand forecasting, multivariate seasonality
• Weather: multivariate long-horizon forecasting
• Traffic: periodicity and weekday/weekend behavior
• Finance / multi-asset: gold, silver, oil, Nasdaq, and FX volatility, regime, and range forecasting experiments

Seminar Track

The seminar/ directory contains a five-day financial time-series forecasting course:

1. Financial time-series problem framing
2. Data analysis, labels, and leakage
3. Baselines and statistical models
4. ML, deep learning, and foundation models
5. Backtesting, risk, and research memo writing

This track intentionally avoids general finance coverage and focuses on timestamp-correct, code-first financial forecasting workflows across gold, silver, oil, Nasdaq, and FX.

Quick Start

just install
just test
just benchmark

For notebook-based exploration:

just notebook

Guiding Principle

Baseline first. Problem definition first. Production constraints first.

A forecasting model is only useful when its prediction target, time horizon, validation method, and business decision are clearly defined.

References

See papers/README.md for the paper reading map.