Factor models convert a portfolio from a list of positions into a map of exposures. The purpose is not to explain away performance. It is to separate rewarded systematic risk, unintended tilts and residual behaviour that may actually deserve attention.

01 / From returns to exposures

A simple linear factor model writes excess return as r_i,t - r_f,t = beta_i' f_t + epsilon_i,t. The factor vector can include market, size, value, profitability, investment, momentum, sectors, rates, credit, volatility or custom signals. The residual is what remains after the chosen risk language has spoken.

02 / Alpha versus hidden beta

Many strategies look exceptional until their returns are projected onto the right factors. A portfolio can appear diversified at position level while secretly concentrating in one macro sensitivity, one liquidity premium or one crowded style. Factor decomposition is a way to make those implicit bets explicit.

03 / Stability matters more than one regression

Static beta estimates are fragile. Rolling windows, expanding windows and regime-specific regressions often tell a richer story: exposures drift, factors become correlated, residual variance changes and the same strategy can behave differently in crisis periods. The model should surface that instability instead of smoothing it away.

04 / Data and construction risk

Factor results depend on universe definition, weighting, rebalance timing, currency, region, survivorship bias and frequency. A clean regression on mismatched factors can still be wrong. In practice, the first question is not the t-stat. It is whether the factor library actually represents the portfolio being tested.

05 / Dashboard logic

A factor interface should show current exposures, contribution to risk, rolling beta bands, residual drawdowns and factor crowding. The useful output is not a table of coefficients. It is a warning system for hidden concentration and regime-dependent behaviour.

A factor model is a translation layer: it turns performance into exposure, and exposure into decisions.

Reference note: Kenneth French Data Library.