Factor Models Construction Process

This project aims to build a stock scoring and backtesting system based on classical factor investing techniques. The goal is to start simple, generate interpretable results, and gradually expand into more sophisticated methods.

Project Settings

• Universe: All historical constituents of CSI 300, CSI 500, and CSI 1000.
  • Stock Universe Document
• Date Range: 2000-01-01 to 2025-12-31
• Frequency: Monthly
• Data Source: Tushare via bagel-tushare (local MySQL database)
• Strategy Type: Long-only, holding top 20–50 stocks

Factor Library

The initial factor set will cover standard categories (all oriented so that higher = better):

• Market
• Size
• Value
• Momentum
• Quality
• Investment
• Volume

A detailed factor dictionary (definitions, formulas, and data fields) will be maintained in a separate file.

Factor Evaluation

Each factor will be tested individually using:

• Information Coefficient (IC) > 0.05
• Rank IC > 0.5
• Quantile Analysis: IC across quantile portfolios
• Correlation Check: pairwise correlation < 0.8 (to reduce redundancy)

Validation Methodology (Walk-Forward)

To better mimic real-time investing, I will use walk-forward validation rather than a one-off split.

1. Timeline & windows
   • Expanding window: at date $t$, fit preprocessing/weights using all data up to $t$, then test on $t \to t+1$.
   • Optionally compare with a rolling window (e.g., last 60 months) to check regime sensitivity.
2. At each rebalance date $t$
   • Universe: only stocks investable at $t$.
   • Preprocessing (fit on train only):
     • Winsorize (1%/99%), z-score by date, optional size/industry neutralization.
   • IC testing: compute cross-sectional IC/RankIC between factor at $t$ and future return $t \to t+1$.
   • Scoring & portfolio: form factor scores using equal/IC/ICIR weights calibrated on the train window only.
3. Aggregation & inference (OOS only)
   • Time series of OOS ${\text{IC}_t}$. Report mean IC, ICIR, hit rate, and HAC (Newey–West) t-stat.
   • Portfolio OOS performance: CAGR, vol, Sharpe, maxDD, turnover, and cost-adjusted returns.
4. Acceptance bar (OOS, monthly)
   • Mean RankIC > 0 with HAC t-stat ≥ 2
   • ICIR ≥ 0.3
   • Hit rate ≥ 55%
   • Factor correlation with others < 0.8
5. Practical cautions
   • Avoid overlapping returns for multi-month horizons (or use larger HAC lags).
   • Keep preprocessing parameters fixed within each train window.
   • Always log data availability dates to prevent look-ahead.

Stock Scoring Framework

Scores will be calculated as a linear combination of selected factors. Initial methods to test:

1. Equal weights
2. IC-weighted
3. ICIR-weighted

Portfolio Construction

• Rebalance: Score at month-end, trade at next month’s open
• Holding Sizes: Test portfolios of 20, 30, 40, and 50 stocks
• Weighting: Equal-weighted portfolios to start
• Transaction Costs: Apply simple cost models (e.g., 10–20 bps per trade side) to adjust returns. Report turnover, cost-adjusted performance, and sensitivity to different cost assumptions.

Future improvements: test alternative rebalancing frequencies, sector neutrality, and more sophisticated transaction cost modeling (e.g., volume-based slippage or market impact).

Roadmap

1. Phase 1: Build factor library and run baseline equal-weight backtests.
2. Phase 2: Add IC/ICIR weighting, factor correlation pruning, and walk-forward validation.
3. Phase 3: Expand portfolio construction with turnover control, transaction costs, and risk adjustments.

✅ Deliverable: For each rebalance date, generate factor scores, stock rankings, and portfolio backtest results — published as a series of blog posts tracking progress.