[MLP Project 4] Hull Tactical – Market Prediction under Volatility Constraints

📌 1. Project Overview

Detail	Description
Course	CS 53744 Machine Learning Project
Task	Time-series regression to predict daily excess returns of the S&P 500 and design a volatility-constrained allocation strategy.
Dataset	Kaggle Competition – Hull Tactical: Market Prediction
Goal	(1) Predict market_forward_excess_returns, (2) map predictions to daily weights w ∈ [0, 2], (3) satisfy σ_strategy ≤ 1.2 × σ_benchmark, (4) maximize a Modified Sharpe ratio.
Evaluation Metric	Kaggle: Modified Sharpe ratio. Local: OOF RMSE & Correlation (for prediction quality) + Sharpe, volatility ratio, cumulative return (for strategy performance).
Final Model	ElasticNet (PCA features) + LightGBM (raw FE) blended (0.95 / 0.05) + volatility-constrained allocation strategy
Baseline Models	Mean-prediction baseline, standalone ElasticNet, standalone LightGBM, standalone XGBoost (tested, excluded from final blend)
Key Insight	No model significantly beats the baseline in RMSE (consistent with EMH), but a carefully regularized blend + mild leverage (k = 0.5) achieves a small Sharpe improvement under 120% volatility cap.

---

👥 2. Team Information

Role	Name	GitHub ID
Member	박원규	@keiro23
Member	이유정	@yousrchive
Member	정승환	@whan0767

---

🏆 3. Final Performance Summary

The final pipeline consists of:

1. Rich feature engineering (lags, rolling stats, regimes, macro shocks, interactions)
2. Time-series cross-validation (walk-forward) without leakage
3. Model comparison & blending (ElasticNet + LightGBM)
4. Strategy evaluation under a volatility constraint (≤ 120% of S&P 500)

3.1 Prediction Performance (OOF, Time-Series CV)

Using 5-fold TimeSeriesSplit on the feature-engineered train set:

Model	RMSE (mean ± std)	Corr (mean ± std)	Comment
Baseline	≈ 0.0108 ± 0.0027	≈ 0.00	Train-mean prediction
ElasticNet	≈ 0.0111 ± 0.0028	≈ 0.03–0.04	PCA(15) + ElasticNet
LightGBM	≈ 0.0122 ± 0.0025	≈ 0.02–0.03	Raw FE, tree-based boosting
XGBoost	≈ 0.0124 ± 0.0025	≈ 0.03–0.04	Slightly worse than ElasticNet / LightGBM
Blend (EN+LGB)	≈ 0.0115	≈ 0.035	0.95 ElasticNet + 0.05 LightGBM (RMSE-opt.)

→ Takeaway: No single model clearly dominates the baseline; any predictability is extremely weak, consistent with EMH.

3.2 Strategy Performance (Vol-Constrained Allocation)

We convert blended predictions to daily weights:

• Standardize blended prediction: z_t
• Define weights: w_t = clip(1 + k·z_t, 0, 2)
• Search k ∈ [0, 50] with step 0.5 under constraint σ_strategy ≤ 1.2 × σ_benchmark

Best k (under constraint): k = 0.5

Metric	Benchmark (w = 1)	Blend Strategy (k = 0.5)
Mean daily excess return	≈ 0.000265	≈ 0.000331
Volatility ratio	1.0	≈ 1.20 (capped)
Annualized Sharpe	≈ 0.378	≈ 0.393
Final cumulative return	≈ 0.400	≈ 0.491

Interpretation: The performance gap is small, but under a strict volatility cap it indicates that weak yet non-zero structure in the feature space can be translated into a slight Sharpe improvement, which is conceptually consistent with EMH’s “very limited predictability” view.

---

⚙️ 4. How to Reproduce Results

We separate the workflow into:

1. Local / offline pipeline (EDA, FE, TS-CV, backtesting, model export)
2. Kaggle online inference (evaluation API using predict(test: pl.DataFrame))

4.1. Environment Setup & Dependencies

1. Create & activate a virtual environment (local):

python -m venv .venv
source .venv/bin/activate  # Linux/macOS
# .\.venv\Scripts\activate  # Windows

2. Install required packages:

pip install -r requirements.txt

Key libraries:

• numpy, pandas, polars
• scikit-learn
• lightgbm, xgboost (optional)
• matplotlib

---

4.2. Data Preparation

1. Download train.csv and test.csv from the Hull Tactical Kaggle competition.
2. Place them inside the data/ directory at the project root:

Project4/
└── data/
    ├── train.csv
    └── test.csv  # mock test for structure reference

---

4.3. Local Pipeline: Feature Engineering, TS-CV & Backtesting

The offline pipeline does three things:

1. Feature engineering on train (generate_FE_interaction_regime)
2. Time-series CV + OOF predictions (ts_cv_oof_predictions)
3. Blend optimization + strategy backtest (search_best_k_for_blend)

---

4.4. Kaggle Inference: Online predict() with Evaluation API

The final result earned from the local backtest is applied to the submission code in src/elastic-lgmb_weight_scale.ipynb.

Kaggle’s evaluation environment:

• No internet

• You receive test batches with:

  • Features M*, E*, I*, P*, V*, S*, MOM*, D*
  • lagged_forward_returns, lagged_risk_free_rate, lagged_market_forward_excess_returns

• You must implement:

def predict(test: pl.DataFrame) -> float:
    ...

Core idea:

• Maintain a buffer of past rows in memory
• Use lagged_market_forward_excess_returns as the y_{t−1} equivalent
• Reproduce the same FE logic as train, but online & incremental
• Select feature_list columns in the correct order
• Apply models and blend predictions as trained locally.

You then typically plug this into the provided evaluation template (default_inference_server) and submit your solution. This process is often demonstrated within a Kaggle-specific notebook.

---

📁 5. Project Directory Structure

Project4/
├── data/
│   ├── submission.csv
│   ├── test.csv
│   └── train.csv
│
├── notebook/
│   ├── 01_eda_baseline.ipynb
│   ├── 02_Feature_Engineering_PCA.ipynb
│   └── 03_Modeling_with_Backtest.ipynb
│
├── src/
│   ├── baseline_submission.ipynb
│   ├── elastic-lgmb_weight_scale.ipynb
│   └── submission.csv
│
├── requirements.txt
└── README.md

---

🧩 6. Notes & Alignment with Course Requirements

• Baseline vs Improved Models

  • Baseline: mean predictor
  • Improved: ElasticNet, LightGBM, blended model

• Feature Engineering & Validation Strategy

  • Rich FE on lagged targets and macro variables
  • TimeSeriesSplit walk-forward CV to avoid leakage

• Local Sharpe-variant & Volatility Plots

  • Backtesting code generates cumulative return and volatility ratio plots

• Kaggle Leaderboard

  • Final submission created via the Kaggle inference pipeline above
  • Screenshot and commentary included in the PDF report

• EMH Discussion

  • Report links small Sharpe improvement and weak predictability back to EMH (approx. weak-form consistency).