AI-Driven Dynamic Portfolio Optimizer (TSFM Edition)

English | 한국어 (README.ko.md) | 日本語 (README.ja.md)

Zero-shot time-series foundation models (TimesFM 1.0, Amazon Chronos-2) forecast expected returns; a Markowitz-style QP (cvxpy) builds long-only portfolios with sector and name caps. A two-tab Gradio app exposes single-asset forecasts and multi-asset optimization.

For the full narrative—evaluation rubric, model cards, notebook metrics, and gap analysis—see REPORT.md (long-form; not duplicated here).

Demo Preview

Features

• Data: Automated multi-source pipeline (Kaggle S&P 500 bulk, yfinance, FRED + Kaggle macro) → data/sp500_macro_master.csv
• Forecasting: Dual-model weighted ensemble (Chronos-2 with macro covariates + univariate TimesFM 1.0)
• Optimization: Sharpe-oriented QP with budget, long-only, GICS sector ≤30%, single-name ≤25%
• UI: Gradio + Plotly (app.py): forecast tab + portfolio tab

Tech stack

torch, transformers, timesfm, chronos-forecasting, pandas, numpy, yfinance, fredapi, cvxpy, gradio, plotly

---

Quick start

Requirements: Python 3.10+ recommended; CUDA GPU strongly recommended for Chronos-2 / TimesFM inference.

python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

Configure credentials (next section), then build data and run the app.

Build the master dataset (see scripts/build_dataset.py for flags and behavior):

python scripts/build_dataset.py

Launch the dashboard:

python app.py

Optional: python scripts/preload_models.py to warm-cache Hub weights; python scripts/run_experiments.py for scripted experiments.

---

Environment variables (.env)

1. Copy the template and edit values (never commit real secrets; .env is gitignored):

cp .env.example .env

2. Loading behavior: src/forecast.py calls load_dotenv(), so python app.py picks up .env from the project root automatically. scripts/build_dataset.py reads os.environ only—it does not load .env by itself. Either export variables in your shell, use direnv, or run once in Bash:

set -a && source .env && set +a && python scripts/build_dataset.py

Get tokens from: Hugging Face settings, Kaggle account API, FRED API keys.

---

Notebooks

Rendered figures and HTML exports live under notebooks/ when cells are executed.

Repository layout (high level)

app.py                 # Gradio entrypoint
src/                   # Forecast + optimization modules
scripts/               # Dataset build, experiments, tests
data/                  # Built CSV (raw paths may be gitignored)
notebooks/             # Analysis + outputs
docs/                  # Extra templates / drafts
REPORT.md              # Full project report (long)

License

This project is licensed under either of

• Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.

• Copyright © 2026 Jaehyun Park (see Required Notice in LICENSE). Redistributing the software requires passing along this license (or its URL) and the required notice lines.
• Third-party stack: PyTorch, Hugging Face models, Python packages, and datasets remain under their licenses; you must comply with those upstream terms in addition to this repository’s LICENSE.
• If you are unsure whether your use is noncommercial, or you want a commercial license, contact the copyright holder (see LICENSE).

The authoritative legal text is LICENSE; the above is a non-binding summary only.