Returns & risk
Constructed adjusted returns, arithmetic and geometric annualization, volatility, Sharpe ratios, and volatility drag.
FLAGSHIP QUANTITATIVE RESEARCH · 2026
Equity Volatility Research Pipeline
Strategy backtesting, market-regime analysis, volatility persistence, and supervised forecasting under a leakage-aware out-of-sample protocol.
PURGED HOLDOUT · LOWER IS BETTER
HAR (1, 5, 22-day)9.08%
EWMA (λ = 0.94)9.30%
Random walk10.28%
Test period: May 2013–January 2016
01 / RESEARCH QUESTION
CAN RECENT VOLATILITY STRUCTURE HELP?
Can a multi-horizon model improve forecasts of the next 21-day realized-volatility window?
The project began with returns and a simple trading rule, then followed the evidence toward a more defensible question. Directional returns showed little persistence; the magnitude of returns and realized volatility did.
02 / METHOD
Strategy validation
Tested a 20-day SMA rule with a one-period execution lag, turnover measurement, and costs from 0 to 100 basis points.
Persistence
Measured rolling realized volatility and autocorrelation in returns, absolute returns, and volatility at multiple horizons.
Forecasting
Compared random walk, historical mean, EWMA, and HAR forecasts against a genuinely forward-looking target.
Out-of-sample design
Used a chronological 70/30 split and purged 21 training labels whose target windows reached into the test period.
Executable controls
Encoded methodological safeguards as 15 tests covering timing, leakage, alignment, solver agreement, and reproducibility.
03 / FINDINGS
Stricter validation changed the interpretation of the results.
Look-ahead bias, overlapping evaluation targets, and split-boundary leakage each made a result appear stronger than it was.
STRATEGY ROBUSTNESS · 2007–2016
0 of 4 beat buy-and-hold.
The net long-only SMA rule underperformed buy-and-hold across MSFT, IBM, SBUX, and AAPL. It lost money on three names. That null result matters: the rule did not generalize.
Useful nuanceThe strategy protected capital during parts of the 2008 crisis, then missed large portions of the recovery.
TARGET INTEGRITY
The model ranking reversed.
An overlapping answer key made the random walk look dominant because consecutive 21-day windows shared 20 observations. Using the next non-overlapping window reversed the ranking and made EWMA the best baseline.
PURGED HOLDOUT
HAR retained a modest edge.
The unpurged HAR result was 9.03%. Removing 21 boundary labels that reached into the test period raised RMSE to 9.08%. The corrected figure still beat EWMA by 2.34% and the random walk by 11.70%.
04 / VALIDATION ARCHITECTURE
Research controls are part of the implementation.
The project turns methodological claims into executable checks so a reviewer can test them rather than take them on trust.
05 / REPRODUCIBLE CODE
Two views of the same research code.
The repository preserves the development path while providing a modular, tested implementation for reviewers who want production-style structure.
RESEARCH WALKTHROUGH
Readable development sequence
A single guided script preserves how the analysis progressed from first principles through the final forecasting experiment.
MODULAR PACKAGE
Reusable research modules
Separate modules for data, returns, strategy logic, volatility, forecasting, and figures reproduce the same results.
INTEGRITY SUITE
15 passing tests
Tests verify timing, target construction, purging, alignment, numerical agreement, annualization, and deterministic reproduction.
06 / LIMITS & NEXT RESEARCH
What this result does not claim
- HAR was evaluated on one asset, one split, and one test period.
- Overlapping forward targets reduce the effective number of independent errors.
- Transaction costs are stipulated rather than estimated from order size and liquidity.
- The long-only strategy cannot profit from declines.
- The project demonstrates research design—not production alpha.
What comes next
- Walk-forward evaluation and multiple chronological splits.
- Cross-asset HAR estimation and residual diagnostics.
- Mean absolute error and a Diebold–Mariano comparison.
- GARCH, semivariance, and leverage-effect extensions.
- Realized-versus-implied volatility with suitable options data.