UCLA MSBA Time-Series Foundation Model Lab

Files:

• tsfm_foundation_eval_lab.ipynb: custom foundation model hands-on notebook
• tsfm_chronos_eval_lab.ipynb: Using OSS foundation model (zero-shot + eval + LLM-as-judge)
• requirements.txt: minimal local dependencies

What this lab teaches:

• Foundation-model workflow for time series: pretrain, zero-shot, few-shot, fine-tune
• Proper time-series evaluation: temporal splits, rolling-origin backtests
• Metrics: MAE, RMSE, sMAPE, MASE
• Why baseline comparisons matter
• Why fine-tuning does not automatically improve results

OSS usage focus:

• Zero-shot forecasting workflow (Chronos path if available; offline fallback if not)
• Metrics + baselines + comparison discipline
• LLM-as-judge for explanation quality (Ollama path if available; fallback heuristic if not)
• Reflection on deployment readiness and monitoring
• Includes a 30-minute student exercise (embedded in the Chronos lab): The Bruins Cold Brew Challenge

Run locally:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
jupyter notebook tsfm_foundation_eval_lab.ipynb

For the OSS zero-shot + eval + LLM-as-judge path, open tsfm_chronos_eval_lab.ipynb instead.

Notes:

• The core notebook is CPU-friendly and runs with numpy + pandas.
• matplotlib is optional (plots are skipped if unavailable).
• The Chronos notebook includes an optional section to plug in a real TSFM (Chronos/TimesFM) while reusing the same evaluation harness.

---

Standalone challenge: Bruins Cold Brew

• bruins_cold_brew_challenge.ipynb — Standalone student-facing forecasting challenge (30 min).
• bruins_cold_brew_demand.csv — Dataset for the challenge (daily demand + context variables: date, avg_temp_c, rain_mm, campus_event, promo_flag, exam_period, holiday_break, is_weekend, cups_sold, split).

Exercise objectives

Students will:

1. Run a zero-shot forecast using a foundation-model style forecaster.
2. Compare it against strong baselines (e.g. naive, seasonal).
3. Tune or calibrate the forecast using recent validation data.
4. Re-evaluate on a held-out test period.
5. Make a deployment recommendation (whether the workflow is trustworthy for staffing and inventory).