Research design

Methodology

A dynamic commodity universe, five signal families, and a Friday-to-next-period portfolio timeline.

Dynamic Universe

The universe is built at the product level from domestic futures contracts. A product must meet the listing-age rule and a 20-day rolling liquidity threshold based on the combined volume of the daily main and sub-main contracts. The main/sub-main approximation selects the two real contracts with the highest volume on each date. Friday observations are used as the weekly rebalancing cross-section.

Factor Families

Five families are implemented: term structure, momentum/trend, warehouse receipts (inventory), position/open interest, and volatility. Term-structure carry uses:

-(far_price - near_price) / near_price * 365 / (far_maturity_days - near_maturity_days)

The calculation then applies a K-day within-product rolling mean. The spot-main variant sets the spot leg maturity to zero. Momentum includes price momentum, cumulative intraday step, signed Garman-Klass volatility, and cross-sectional daily-return rank components. Inventory measures a smoothed warehouse-receipt change; position compares current open interest with its K-day average. Volatility variants use OHLC-based estimators.

Portfolio Construction And Alignment

On each Friday, valid product signals are sorted high-to-low and allocated to five balanced groups. Group 1 is the highest factor bucket and group 5 the lowest. Membership begins earning returns on the next trading date and is held through the following rebalance date. Daily returns are equally weighted within group; the reported long-short return is 0.5 * (group_1 - group_5). This next-period alignment prevents the rebalance-day close from entering the realized return.

Weekly IC is Spearman correlation between the Friday factor cross-section and subsequent holding-period product returns. The industry-neutral checks preserve the same factor and return timing while evaluating category-adjusted exposures. Multi-factor portfolios first score available factors by either five-bucket 100-point scores or cross-sectional z-scores, then average the available scores before the same quintile backtest.

Known Implementation Limits

This is a replication using currently available local inputs, not a claim of exact vendor-data identity. Main and sub-main contracts are volume-based approximations; spot coverage and some multi-factor components may be incomplete. See limitations for the full publication caveats.