Time-Series Momentum & Carry Factor Library — Chinese Commodity Futures
A research pipeline for two classical factor families on the Chinese commodity futures market, built on roughly 15 years of daily data across 30+ continuous (main-contract) series.
1. Factor Construction
Time-series momentum. Multi-window return signals, volatility-normalised so that contracts with very different realised volatility contribute comparably to the composite.
Carry. Term-structure slope, i.e. the annualised roll yield implied by the spread between nearby and deferred contracts.
Both families are then cross-sectionally standardised and sector-neutralised, so that the surviving signal is not simply a bet on one commodity sector (energy, ferrous, agriculture, …).
2. Evaluation Suite
Rather than reporting a single headline Sharpe ratio, every factor is passed through a standard diagnostic battery:
| Diagnostic | What it answers |
|---|---|
| IC / ICIR | Is the rank correlation with forward returns positive, and is it stable relative to its own volatility? |
| Quintile monotonicity | Does the return ordering hold across all five buckets, or is the signal driven by one extreme tail? |
| Half-life | How fast does the signal decay — and therefore how much turnover does it demand? |
3. Results
| Portfolio | Annualised Sharpe | Notes |
|---|---|---|
| Time-series momentum | 1.15 | |
| Carry (long–short) | 0.90 | IC 0.06, strictly monotonic across quintiles |
| IC-IR weighted composite | 1.60 | Max drawdown held within 9% |
The composite weights the two families by their IC-IR, which is the natural weighting when the goal is to maximise the information ratio of the combined signal under an approximate independence assumption.
4. Caveats
The strict quintile monotonicity of the carry factor is the more informative result here — a monotone bucket ordering is much harder to produce by chance than a good Sharpe ratio on a single long–short leg. The composite Sharpe of 1.6 is reported before transaction costs and slippage; on a factor with this half-life, execution assumptions materially affect the realisable number.