Building a UCITS‑Only Quarterly Top‑5 Rank‑Weighted Momentum Strategy

Momentum strategies have long delivered strong, robust risk‑adjusted returns with relative simplicity. In this project, I build on three key studies that directly informed my approach:

• Moskowitz, Ooi, and Pedersen (2012): Documents the prevalence and persistence of time‑series momentum across futures and equity indices.
• Hurst, Ooi, and Pedersen (2014): Explores cross‑sectional and time‑series momentum in global equity markets.
• Krauss, Do, and Huck (2018): Compares simple momentum rules against machine‑learning models, finding basic look-back strategies difficult to beat.

Findings guided my design of a UCITS‑compliant ETF rotation strategy.

Strategy Methodology & Backtesting Framework

• Universe: UCITS‑listed, total‑return ETFs accessible via my Interactive Brokers retail account. This limitation forced a narrower ETF set — excluding some high‑quality non‑UCITS funds.
• Data Range: May 2015 to May 2025 daily NAVs.
• Momentum Signal: 12‑month total return look-back; alternate windows (6, 9, 24 m) were explored in experimentation.
• Selection & Weighting: At each quarter‑end:
  1. Rank ETFs by momentum.
  2. Select top five.
  3. Allocate capital proportionally to rank (highest momentum = highest weight).
• Trading Costs & Constraints: Strategy assumes 1 % p.a. annualized flat trading cost; benchmarks (ACWD, S&P 500) assume 0.1 % p.a. turnover. Position limits and liquidity filters ensure UCITS compliance.

Backtests were conducted in a modular Python pipeline, with separate data ingestion, signal generation, portfolio construction, and performance modules. This structure facilitated iterative experiments but grew complex over time—reinforcing code modularization’s importance.

Full methodology details: docs/methodology.md

Parameter Search & Experimental Design

To optimize performance without overfitting, I employed:

• Bayesian Optimization using the scikit-optimize library to tune look-back windows (6–24 m) and weight scaling parameters. This approach efficiently explored parameter space rather than brute‑force grid search.
• Enhancements Tested:
  • Volatility scaling (inverse 3‑month volatility).
  • Regime filters (50 d vs. 200 d moving average crossover).
  • Weighting schemes (raw returns vs. rank‑based scores).

While Bayesian optimization yielded strong in‑sample metrics, these tuned parameters produced unrealistically high backtest returns, indicating overfitting.

See full experiments: docs/experiments.md

Walk‑Forward Validation & Overfitting Mitigation

To ensure out‑of‑sample robustness, I used a walk‑forward optimization (WFO) scheme:

1. Training Window: 2 years.
2. Testing Window: 1 year.
3. Step Size: 6 months.

At each fold, parameters (fixed 12 m look-back, rank weighting) were applied without re-optimizing to avoid introducing look‑ahead bias. In 14 out‑of‑sample folds:

The consistent performance and controlled drawdowns confirmed that the simple 12 m→3 m top‑5 rotation was most robust.

Detailed walk‑forward results: docs/results.md

Performance Summary (May 2015–May 2025)

Highlights: lower drawdowns, positive alpha at ~0.78 beta exposure, and robust out‑of‑sample results.

Practical Learnings & Future Directions

1. Correct Backtesting Matters | Walk‑Forward Importance: Parameter tuning via Bayesian optimization led to impressive in‑sample gains but failed out‑of‑sample. WFO exposed overfitting, guiding me back to the simple 12 m setup.
2. Universe Constraints: Retail account limitations (IBKR UCITS ETFs) significantly cap performance. Non‑UCITS ETFs improved returns in parallel tests—highlighting data scope’s impact.
3. Code Modularization: As experiments grew, clean, well‑documented modules ensured reproducibility and easier debugging.

Read more: GitHub: momentum-strategy

Get in touch:

• 📧 alexeyg377@gmail.com
• 🔗 LinkedIn Profile