A rigorous, end-to-end quantitative factor research framework built in Python. Demonstrates the full research workflow: hypothesis formation, data engineering, factor construction, backtesting, and honest interpretation of results.
quant_factor_research/
├── main.py # Orchestrator — run this
├── requirements.txt
├── src/
│ ├── pipeline.py # Section 1: Data download and cleaning
│ ├── factors.py # Section 2: Factor signal construction
│ ├── value.py # Section 2a: Value factor (earnings yield)
│ ├── backtest.py # Section 3: Portfolio formation and performance metrics
│ └── analysis.py # Section 4: Charts and summary tables
├── data/ # Auto-created — cached parquet files
└── output/ # Auto-created — all charts saved here
pip install -r requirements.txt
python main.pyOptions:
python main.py --start 2010-01-01 --end 2023-12-31 # custom date range
python main.py --refresh # force re-download data
python main.py --no-cost-adjust # skip transaction cost nettingThe first run downloads ~500 stocks × 20 years of data. Expect 5–10 minutes. Subsequent runs load from the parquet cache and finish in under a minute.
- Construction: Compound return from month t−12 to t−2 (12-1 convention). The most recent month is skipped to avoid short-term reversal contamination.
- Academic basis: Jegadeesh & Titman (1993) — Returns to Buying Winners and Selling Losers. One of the most replicated findings in empirical finance.
- Hypothesis: Stocks with strong past performance continue to outperform over the subsequent 3–12 month horizon.
- Portfolio: Long Q5 (highest past return), short Q1 (lowest).
- Construction: Trailing 252-day realized standard deviation of daily returns, annualized. Signal is negated so high signal = low volatility.
- Academic basis: Ang et al. (2006) — The Cross-Section of Volatility and Expected Returns. The "low vol anomaly": less volatile stocks earn higher risk-adjusted returns than theory predicts.
- Hypothesis: Low-volatility stocks are systematically underpriced because institutional investors with leverage constraints prefer high-vol names.
- Portfolio: Long Q5 (lowest volatility), short Q1 (highest volatility).
- Construction: Earnings Yield = trailing twelve-month Net Income / Market Cap. Net Income sourced from yfinance annual income statements. Availability date set to fiscal year end + 60 days to approximate reporting lag. Market cap computed as current shares outstanding × end-of-month price.
- Academic basis: Fama & French (1992) — The Cross-Section of Expected Stock Returns. Earnings yield is one of the strongest documented value signals.
- Hypothesis: Stocks trading cheaply relative to their earnings are underpriced by the market and expected to outperform over a 1–3 year horizon.
- Portfolio: Long Q5 (highest earnings yield), short Q1 (lowest).
- Limitations: Current share count used in place of historical; reporting lag
is a proxy, not an actual announcement date. See
src/value.pyfor full details.
- Construction: Prior 1-month return, negated.
- Academic basis: Jegadeesh (1990) — Evidence of Predictable Behavior of Security Returns. Strong winners over one month tend to give back gains.
- Hypothesis: Market microstructure effects and liquidity provision dynamics cause mean reversion at the 1-month horizon.
- Portfolio: Long Q5 (prior month's losers), short Q1 (prior month's winners).
- Universe: S&P 500 current constituents (see Limitations).
- Rebalancing: monthly, at end-of-month close.
- Portfolio construction: equal-weighted within each quintile.
- Signal → return shift: signal computed at t, return measured at t+1. This is the core lookahead bias prevention mechanism.
| Metric | Description |
|---|---|
| Annualized Return | CAGR from monthly returns |
| Sharpe Ratio | Ann. return / ann. vol (RF = 0%) |
| Max Drawdown | Largest peak-to-trough loss |
| Hit Rate | % of months with positive return |
| Avg Turnover | Avg % of portfolio replaced each month |
- Default: 10 bps one-way (20 bps round-trip) per rebalance.
- Applied proportionally to monthly turnover of each quintile.
- Conservative for liquid large-cap stocks; adjust for real-world conditions.
For each factor:
{Factor}_cumulative_returns.png— All quintiles + L/S gross vs net{Factor}_quintile_bar.png— Annualized return per quintile (monotonicity test){Factor}_drawdown.png— Underwater equity curve{Factor}_rolling_sharpe.png— 36-month rolling Sharpe ratio
Combined:
factor_correlation_heatmap.png— Pairwise L/S return correlationsall_factors_combined.png— All factors' L/S returns on one chart
| Factor | Ann. Return | Sharpe | Max Drawdown | Hit Rate | Avg Turnover |
|---|---|---|---|---|---|
| Momentum | −4.75% | −0.20 | −68.69% | 52.86% | 24.5% |
| Low Volatility | −13.70% | −0.62 | −94.19% | 40.87% | 7.3% |
| Reversal | +1.36% | +0.17 | −27.86% | 47.90% | 78.4% |
Long-short portfolios, gross of transaction costs. Period: 2005–2024.
Result: The long-short portfolio returned −4.75% annualized with a Sharpe of −0.20. Rather than the expected pattern of Q5 (past winners) outperforming Q1 (past losers), the quintile returns ran almost perfectly in reverse — Q1 earned 16.45% annualized while Q5 earned only 14.15%.
Why this happened: This is a direct consequence of survivorship bias. In the academic literature, momentum is documented on point-in-time databases where past losers include stocks that subsequently went bankrupt or were delisted — the true tail of the distribution. In this dataset, the only "past losers" available are stocks that remained in the S&P 500 through to today, meaning they recovered by definition. The genuine losers — companies that failed — were silently excluded from the universe before the backtest even began. What looks like a contrarian signal is actually a sample of resilient survivors.
This result is consistent with Israel & Moskowitz (2013), who show that momentum profits are highly sensitive to the treatment of delisted stocks, and that survivorship-biased samples can substantially distort or reverse the signal.
Result: The long-short portfolio returned −13.70% annualized with a Sharpe of −0.62 and a maximum drawdown of −94.19% — the worst result of the three factors by a wide margin. Q1 (highest volatility stocks) earned 22.51% annualized while Q5 (lowest volatility) earned only 10.98%.
Why this happened: The low-vol anomaly is the most survivorship-sensitive of the three factors tested. High-volatility stocks have a bimodal outcome distribution: they either blow up (bankruptcy, delisting) or produce outsized returns. A point-in-time database captures both outcomes. A survivorship-biased database captures only the winners — exactly the high-vol stocks that happened to survive and thrive. The result is an artificial "high volatility premium" that completely inverts the academic finding.
In the Ang et al. (2006) paper, the low-vol anomaly was documented using CRSP data with delisted returns included. The contrast between their findings and the results here illustrates why data quality is the most critical input in factor research — not the sophistication of the model.
Result: The long-short portfolio returned +1.36% annualized with a Sharpe of +0.17. This is the only factor that produced a positive long-short return, and the only one directionally consistent with academic predictions.
Why it partially survived: The reversal signal is less affected by survivorship bias than momentum or low-vol because it operates on a 1-month horizon. The bias introduced by excluding failed companies accumulates over longer windows; at one month, the distortion is smaller. The signal is directionally correct — prior losers did modestly outperform prior winners on average.
Why it is not tradeable: The 78.4% average monthly turnover makes this strategy extremely expensive in practice. After realistic transaction costs, the 1.36% gross return would be largely or entirely consumed. The Sharpe of 0.17 is below any reasonable threshold for capital allocation. Academic reversal strategies typically require sub-millisecond execution and near-zero transaction costs to be viable — conditions not available to most investors.
The most important conclusion from this study is methodological, not financial: survivorship bias is large enough to completely reverse two of the most well-documented factors in academic finance. Momentum and low volatility both inverted, while the one factor least sensitive to long-horizon bias (reversal) was the only one to produce a directionally correct result.
This has a direct implication for any quantitative research workflow: the choice of data source is a more consequential decision than the choice of factor model, backtesting framework, or performance metric. A sophisticated backtest on a biased dataset produces confidently wrong answers. The correct response is not to find factors that "work" on the available data, but to understand the structure of the bias and interpret results accordingly — which is what this study attempts to do.
These are documented deliberately — understanding what a backtest cannot tell you is as important as the results themselves.
| Bias | Description | Impact |
|---|---|---|
| Survivorship bias | We use today's S&P 500 list, not point-in-time historical constituents. Companies that were removed due to bankruptcy or poor performance are excluded. | Upward bias on returns, especially for value and momentum strategies. |
| Lookahead in universe | The constituent list itself is known only in hindsight. | Universe selection bias compounds survivorship bias. |
| Price impact / market impact | We assume trades at closing prices with no slippage. Real rebalances move the market. | Overstated returns for high-turnover strategies. |
| Risk-free rate = 0 | Sharpe ratios don't subtract T-bill returns. | Overstated Sharpe ratios in high-rate environments. |
| No fundamental data | Value factor excluded because yfinance book values are not point-in-time safe. Using them would introduce lookahead bias via accounting restatements. | Limits the factor set. |
- Data engineering: batched downloads, caching, missing data handling
- Factor construction: academically grounded, documented assumptions
- Lookahead bias prevention: explicit forward-shift of signals
- Transaction cost modeling: turnover-proportional cost adjustment
- Statistical rigor: Sharpe, drawdown, rolling metrics, monotonicity testing
- Honest reporting: documented limitations, survivorship bias warning
- Jegadeesh, N. & Titman, S. (1993). Returns to Buying Winners and Selling Losers. Journal of Finance, 48(1), 65–91.
- Ang, A., Hodrick, R., Xing, Y., & Zhang, X. (2006). The Cross-Section of Volatility and Expected Returns. Journal of Finance, 61(1), 259–299.
- Jegadeesh, N. (1990). Evidence of Predictable Behavior of Security Returns. Journal of Finance, 45(3), 881–898.
- Fama, E. & French, K. (1993). Common Risk Factors in the Returns on Stocks and Bonds. Journal of Financial Economics, 33(1), 3–56.