Factor Strategies

This is an introductory long-form article for the "Factor Strategy" series: the standard workflow to turn a factor from "idea" to "usable" — the full Alphalens process. It answers the question "Does my factor actually work?" with a replicable checklist.

One-Sentence Conclusion

Stage	What to Do	Entry Point
Build Factor	Start from an explainable idea and write the first version of the factor (see 7-Factor Model for structure)	Beginner
Validate Factor	Alphalens layered backtesting: three tables for IC, turnover, and decay (see complete case of Annualized 36% ORB)	Advanced
Avoid Pitfalls	The three questions on Look-ahead Bias, Survivorship Bias, and Overfitting	Practical

1. Idea First, Factor Second

Factors are not "run out"; they are "thought out." Every factor that survives out-of-sample testing is backed by an economic intuition:

• Momentum: Winners keep winning (Behavioral finance: underreaction, disposition effect)
• Value: Cheap stocks offer compensation (Risk compensation or mispricing)
• Quality/Low Volatility: The market underprices "boring good companies"
• Event-Driven: Predictable drifts from earnings reports, buybacks, and dividends

Before writing code, write one sentence: "I believe stocks of type ___ will outperform because ___." If you can't write this sentence, don't write code. The 7-Factor Model: Besides Size, Market, Momentum, and Value, What Else? is a great structural reference — look at the "shape" of other factors in factor libraries before designing your own.

2. The Alphalens Full Process (Standard Workflow)

The core problem Alphalens solves is: aligning "factor values" with "future returns" and then analyzing them in layers. Its output consists of three things:

12345678# Minimal viable workflow (pseudocode; logic is more important than API)
# 1. Alignment: Factor values (available after T-day close) vs. future N-day returns
#    — Note: "Available after T-day close" means usable on T+1; this is the first gate against look-ahead bias.
# 2. Layering: Divide into 5 groups based on factor values (long-short only looks at the top and bottom groups).
# 3. Three Tables:
#    - Layered Cumulative Returns: Is the top group stably at the top?
#    - IC (Information Coefficient): RankIC mean > 0.03 to warrant a second look.
#    - Turnover and Decay: Factor rebalancing frequency vs. cost; how long is the IC half-life?

Three criteria are indispensable:

1. Layered Monotonicity: The return curves from Q1 to Q5 must be generally monotonic. Factors with crossing or entangled curves are noise, no matter how high the IC.
2. IC Stability: Look at the level via the RankIC mean and stability via ICIR (mean/standard deviation). A factor that only works in bull markets will reveal its flaws through ICIR.
3. Tradability: Multiply turnover by two-sided costs (A-shares start at 2.5 bps). What remains after deduction? Many "30% annualized" factors drop to single digits after costs.

The ORB Strategy with 36% Annualized Return is a complete case from idea to validation; it is recommended to read it alongside this article.

3. Multi-Factor Portfolios: From 1 to N

Once a single factor is usable, the next step is portfolio construction. The 7-Factor Model discusses supplementary dimensions beyond size, market, momentum, and value. When combining factors, remember three rules:

• Orthogonalize First: If the correlation between a new factor and existing factors exceeds 0.7, ask yourself: "Is this bringing new information?"
• Check Marginal Contribution: Add factors one by one and observe the incremental improvement in portfolio Sharpe ratio, rather than each factor's individual performance.
• Control Quantity: There are limited independent risk sources effective in A-shares. After 7±2 factors, most additions are likely noise.

The ESG Score Long-Short Strategy includes a universal code for layered backtesting that you can directly apply to your own factors.

4. The Three Questions (Must-Answer Before Launch)

• Look-ahead Bias: Does every piece of data used in the factor truly exist "at that time"? Use announcement dates for earnings reports, not reporting periods; use daily factors for adjustments.
• Survivorship Bias: Are delisted and ST stocks included in the sample? Looking only at stocks that survived to the present artificially inflates backtest results.
• Overfitting: Were the parameters "selected" or "tuned"? Beautiful curves found after tuning 20 parameter sets are likely to fail out-of-sample.

5. Further Reading

• 7-Factor Model — The shape of factor libraries
• ORB Strategy with 36% Annualized Return — Complete validation case
• ESG Score Long-Short Strategy (with Universal Layered Backtesting Code)
• Genetic Algorithms for Strategy — Search ideas for factor combinations
• The 42nd Factor: 17.6% Annualized — Deep dive into a single factor
• Such a Simple Alpha Calculation Was Actually Wrong — Details determine survival

There is no silver bullet in factor research, only one sentence, three criteria, and three questions. Once you run this workflow smoothly, your factors can truly be considered "verified."