🚀 Forecasting S&P 500 Quarterly Revenue with TFT

Reproducible deep learning pipeline for firm-level quarterly revenue forecasting using the Temporal Fusion Transformer (TFT) on a long-horizon S&P 500 panel (1995Q1–2025Q2).

This project benchmarks TFT against LSTM and classical time series models (ARIMA/SARIMA) under strict chronological splits and leakage-free design.

🔥 Key Results

• Next-quarter (h=1) forecasting: ~9.3% MAPE on 155 S&P 500 firms.
• Strict evaluation:Fully leakage-free pipeline;Chronological train/val/test splits.
• Model comparison:TFT (panel model) vs LSTM vs ARIMA/SARIMA (per-ticker).
• Interpretability:Feature importance via TFT variable selection & attention

🧠 Methodology Overview

• Raw Financial Data
• Feature Engineering (lags, growth, fundamentals, calendar)
• TFT / LSTM / ARIMA Models
• Multi-horizon Forecasting (h = 1, ..., H)
• Evaluation (MAPE, MAE, RMSE).

📊 Dataset

• Universe: 155 continuously listed S&P 500 firms
• Period: 1995Q1 – 2025Q2
• Frequency: Quarterly
• Target: Revenue (log-transformed during training)

⚠️ Data is not included due to licensing. Please place your dataset under data/ following the expected structure.

📌 Highlights

• Unified chronological split (70–15–15).
• Leakage-safe feature engineering.
• Panel modeling with TFT.
• Per-ticker classical baselines.
• Fully reproducible pipeline (fixed seeds, saved configs)

📂 Repository Structure. \

├─ src/ # Python modules (dataset, models, training, utils)
├─ notebooks/ # EDA and experiment runs
├─ figures/ # Plots exported to paper
├─ results/ # Metrics, tables, predictions (small text files)
├─ data/ # (gitignored) raw/processed data, per-ticker CSVs
├─ scripts/ # (gitignored) checkpoints (TFT/LSTM)
├─ requirements.txt # Python deps (or use environment.yml)
├─ .gitignore # Python/Jupyter/LaTeX/data artifacts
└─ README.md

🔧 Setup

Option A — venv (Windows/macOS/Linux)

python -m venv .venv

• Windows: .venv\Scripts\activate
• macOS/Linux: source .venv/bin/activate

pip install --upgrade pip

pip install -r requirements.txt

▶️ Quick Start

• Prepare data python scripts/prepare_data.py
  --input_dir data/raw
  --output_dir data/processed
  --split_scheme 70-15-15
  --seed 2025

• Train TFT python scripts/train_tft.py
  --data_dir data/processed
  --save_dir models/tft_baseline
  --max_epochs 50
  --batch_size 256
  --hidden_size 128
  --lr 3e-4
  --seed 2025

• Evaluate & export tables

• python scripts/eval.py
  --data_dir data/processed
  --model_dir models/tft_baseline
  --out_dir results/tft_baseline

• Train baselines
  --LSTM
  python scripts/train_lstm.py --data_dir data/processed --save_dir models/lstm_baseline

  --ARIMA/SARIMA (per-ticker)
  python scripts/train_arima.py --data_dir data/processed --out_dir results/arima
  python scripts/train_sarima.py --data_dir data/processed --out_dir results/sarima

🧱 Feature Taxonomy

• Static s_i: ticker/CIK (group id), GICS sector (embedded).
• Observed–past o_{i,t}: revenue lags, growth/volatility, rolling stats, lagged fundamentals.
• Known–future k_{i,t+h}: calendar (year/quarter), lag1 assets/equity treated as deterministic at t.
• Exact columns and encodings are defined in src/data/schema.py (to be finalized).

📊 Metrics & Reporting

• Primary metric: MAPE (%); secondary: MAE, RMSE, MdAPE, and accuracy=100−MAPE.

🔬 Experiments

• Ablations: remove groups of features (e.g., drop totalAssets), compare deltas.
• Robustness: sector-wise breakdown, error distribution, sensitivity to horizon/split.
• Interpretability: TFT variable selection / attention diagnostics.
• Reproduce any experiment with a single YAML config (see configs/) and:
• python scripts/run_experiment.py --config configs/tft_baseline.yml

🛡️ Reproducibility

• Fixed seeds, deterministic cuDNN where possible.
• Frozen train/val/test indices saved under data/processed/splits/.
• Environment export:
• pip freeze > requirements-lock.txt

📜 License

• Code: MIT (recommended) — see LICENSE.
• Data: Not included (subject to original provider terms).

📣 Citation

If you use this repo, please cite:

@misc{tft_sp500_quarterly_revenue_2025, title = {Forecasting S&P 500 Quarterly Revenue with Temporal Fusion Transformer}, author = {Wu, Qiping and Collaborators}, year = {2025}, howpublished = {\url{https://github.com/Rockefarmer/Forecasting_SP500_Quarterly_Revenue_with_TFT}} }

🙌 Acknowledgments

• Lim et al., Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting.
• PyTorch
• statsmodels
• Open-source community

💡 About This Project

This repository is part of a Master’s thesis focused on:

• Financial time-series forecasting
• Deep learning (TFT)
• Multimodal extensions (future work with NLP: FinBERT, Llama).