Why a factor model
Factor models are used across investment management to monitor risk and to construct portfolios. One motivation is dimensional: computing the risk of a book with a large number of assets is tractable once the assets are expressed through a much smaller set of common drivers. The second is that the reduction is not merely arithmetic. It puts a risk manager's attention on the broad themes in the book rather than on a matrix of pairwise relationships.
The form
Every asset return is driven by the same set of factors, plus a specific return unique to that asset. Loadings differ by asset and change over time.
Ri(t) = Σj βij Fj(t) + εi(t)
| Term | Meaning |
|---|---|
| Ri(t) | Return of asset i over period t. |
| Fj(t) | Return of factor j over period t. |
| βij | Loading — the sensitivity of asset i to factor j. |
| εi(t) | Specific, or idiosyncratic, return of asset i. |
Assets typically number in the thousands and factors an order of magnitude fewer: an equity model may cover ten thousand names with a factor set of two hundred. Estimation cycles have moved from monthly to weekly to daily. The horizon of the return in the relationship above has to be stated, not assumed.
Estimation
The modeller has to decide what is known and what is to be estimated. That single choice separates the two families of model.
Time series
Each factor return is known for period t. Loadings and specific return are derived through a time series regression, asset by asset.
Implementations include principal component analysis, which yields uncorrelated latent factors and typically explains most of the variance with fewer than thirty of them. Its drawback is serial inconsistency: run day after day, the extracted portfolios can change radically, so vendors condition on the previous day's data to hold the model steady. The alternative — regressing on a set of pre-identified thematic portfolios — gives factors that are easy to name but carries ad hoc choices, and portfolios independent of the estimation technique need not be statistically independent.
Cross-sectional
Loadings are known for each asset. Factor returns and specific returns are derived through a regression across all assets, for each period.
Exposures are built from observable data, normalised, and expressed as z-scores; asset returns are then regressed on them. The factor returns are determined endogenously, by the returns of the assets and the level of exposure. With factor returns in hand it becomes possible to construct factor replicating portfolios, and from those, portfolios tilted towards a factor.
Whichever approach is taken, the output is the same set of objects: exposures, a factor covariance matrix, and factor returns. From those follow portfolio variance, volatility, value at risk, and attribution.
Commodities
Commodities are hard assets spanning agriculture, energy and metals, traded on exchanges since the seventeenth century. The futures contract is what makes them modellable.
| Feature | Consequence |
|---|---|
| Standardisation | The contract standardises what is often a heterogeneous physical product, which introduces delivery optionality. |
| Fixed expirations | Standardised expirations coordinate market behaviour. Crude oil carries monthly maturities out to ten years, more than adequate for most business planning horizons. |
| Two populations | Hedgers and speculators are both present and active in the same contracts. |
| Volatility premium | Commodities show periods of extreme volatility, with evidence of a premium in excess of the compensation for that volatility. |
| Common factors | The literature has established that commodity futures prices are influenced by a small number of common factors — momentum, basis, liquidity, open interest. |
| Correlation structure | The asset class has historically exhibited return relationships that differ from other markets, one reason institutions study it alongside other asset classes. |
The parallels with equities are close. The difference is that the extension of factor models across asset classes has largely bypassed commodities, leaving the coverage that equity managers take for granted unavailable here.
Consequence
01
A book can be sector-neutral and still carry substantial systematic risk. In the published worked example, the long and short legs offset almost exactly across Energy, Metals and Agriculture while the same book ran a basis exposure close to one standard deviation. Sector arithmetic alone would have called it hedged.
02
Representing positions through factor exposures lets a researcher separate common drivers from contract-specific maturity effects.
03
Systematic and specific risk both arise from price fluctuation, but they behave differently. A factor model separates them. In the worked example the two contribute in roughly equal measure to total variance, a split no correlation matrix would surface.
04
Portfolio return can be represented as a linear combination of exposures and factor returns. Applying historical factor returns to a defined set of exposures produces an illustrative stress scenario, not a forecast.
The ARC model
ARC takes the cross-sectional approach. Each contract loads on its sector, its sub-sector, and a set of styles and trading factors. Sub-sectors are nested inside sectors.
The estimation is calibrated so that the two levels agree: risk and factor performance are unchanged whether the model is read at sector or sub-sector level. Aggregation requires no re-estimation.
References
| 1 | Grinold, R. and Kahn, R. (1999). Active Portfolio Management: A Quantitative Approach for Producing Superior Returns and Controlling Risk. McGraw-Hill. |
| 2 | Petram, L. (2014). The World's First Stock Exchange. Columbia Business School Publishing. |
| 3 | Ferris, W. G. (1988). The Grain Traders: The Story of the Chicago Board of Trade. Michigan State University Press. |
| 4 | Sakkas, A. and Tessaromatis, N. (2018). Factor based commodity investing. EDHEC Business School. |
| 5 | Asness, C. S., Moskowitz, T. J. and Pedersen, L. H. (2013). Value and momentum everywhere. The Journal of Finance, 68(3). |