PCA Eigenportfolio Stat Arb (Statistical Factors)

Thesis (edge)

Every risk model needs factors. The usual approach borrows them from economics: market, size, value, sector, momentum. The statistical approach does something different. It asks the data to reveal its own factors.

Take a few hundred stocks and look at how their daily returns move together. Principal component analysis finds the handful of common patterns that explain most of that shared movement. The first pattern almost always turns out to be "everything moves with the market". The next few usually look like sector or industry effects, or a growth versus value tilt, though nobody told the algorithm to look for them.

Once you have those patterns, you can hedge each stock against them. Whatever is left over, the residual, is the part of the stock's move that has nothing to do with any broad pattern in the market. And that residual, historically, tends to drift back toward its own average.

So the trade is: hedge out the common factors, then buy the stocks whose residuals have fallen too far and short the ones whose residuals have risen too far. It is the same instinct as pairs trading, generalised. A pair uses one stock as the hedge. This uses the whole market's shared structure as the hedge.

Where it works (regimes)

• Works well: in stable markets where the correlation structure is persistent, so the components you fitted last month still describe the market this month.
• Works well: in large, liquid universes, because principal components estimated on a small number of names are mostly noise.
• Fails: when the correlation structure shifts abruptly. In a crisis, correlations converge, the first component swallows everything, and the residuals you thought were idiosyncratic turn out to be full of market risk.
• Fails: when a stock has a genuine, company-specific event. The residual moves for a real reason and does not revert. Statistical hedging cannot distinguish "temporary flow" from "the CFO resigned".
• Crowded: this is one of the standard tools of the quant equity industry. Its worst days tend to arrive on everybody else's worst days.

Signals

• The components. Estimate the correlation of returns across the universe on a rolling window and extract the leading components. Each component can be read as a portfolio, which is why they are often called eigenportfolios. Being able to look at one and say "this is basically energy versus tech" is a valuable sanity check.
• How many to keep. This is the central judgement call. Keep too few and real common risk leaks into the residual, so you end up fading genuine factor moves. Keep too many and you hedge away the very idiosyncratic signal you are trying to trade. There is no clean answer, only a defensible rule chosen without looking at future data.
• The residual and its score. Hedge each stock against the retained components, accumulate its residual, and express how far it currently sits from its own typical level. Extreme scores are the trade.
• Reversion speed. Estimate how quickly a stock's residual has historically returned to its average. Names whose residuals revert slowly should be excluded, because you cannot hold long enough to collect.

Portfolio construction

• Neutral to every retained component. After sizing, the book should have effectively no exposure to any factor you extracted. That is what makes it market neutral in a meaningful sense rather than just dollar neutral.
• Wide and shallow: many names, small positions. The signal per stock is weak and only works on average.
• Scale by residual volatility so that the noisiest names do not dominate risk.
• Refit on a schedule, trade continuously. Re-estimating the components every day makes the hedge chase noise and generates pointless turnover. Refit periodically and accept a slightly stale hedge.
• Cost-aware trading: only act on signals large enough to be worth the round trip.

Risk model

• Instability of the factors. This is the defining risk. Statistical factors have no economic anchor. When they change, you have no story to fall back on, and no way to know whether the change is noise or a regime shift until after it has cost you.
• Hidden market exposure. If you kept too few components, or the correlation structure shifted, your supposedly neutral book can carry substantial directional risk exactly when that is most painful.
• Estimation error in the correlation matrix. With hundreds of stocks and a limited history, the estimated correlations are noisy. Shrinkage helps considerably and is close to mandatory here.
• Crowding. Many funds run close variants of this. When they cut risk, residual spreads move against everyone simultaneously.
• Event risk on single names. Mergers, fraud, and delistings do not revert, and they arrive without warning in the name you have the largest residual position in.
• Cost and borrow risk: high turnover, and a short book that can include hard-to-borrow names.

Costs & implementation

• This is engineering-heavy. Rolling correlation estimation, component extraction, exposure calculation, residual tracking, cost-aware optimisation and neutrality constraints all have to work together every day.
• Turnover is high, and the edge per trade is small, so the strategy is unusually sensitive to execution quality. Small improvements in fills matter more than clever modelling.
• Shrinkage on the correlation matrix materially improves stability and should be treated as part of the core method rather than an optional extra.
• Capacity is limited. Residual signals are strongest in the names that are hardest to trade in size.
• Backtesting is easy to get wrong. Extracting components from the full history and then testing on part of it is a subtle and fatal look-ahead. The components must be estimated using only data available at the time.

Failure modes

• Look-ahead in the components. The most common and most damaging error, because it produces a very convincing backtest.
• Choosing the number of components by testing which works best. That is fitting the answer to the data, and the live result will not match.
• Assuming statistical neutrality equals real neutrality. Neutral to your estimated factors is not the same as neutral to the actual risks in the market.
• Ignoring shrinkage. Raw correlation matrices estimated from limited data are unreliable and produce unstable, high-turnover portfolios.
• Trading names with slow-reverting residuals. They tie up capital and deliver nothing.
• Believing the components are permanent. They are a description of the recent past, not a law.

Our Notes & Suggestions

Always look at your factors. Print out the top components as portfolios and see whether they make sense. If the second component is clearly energy versus technology, good, you have found something real. If it is an incoherent mix of unrelated names, it is probably noise, and hedging against noise while trading the leftovers is not a strategy.

Test the stability of the components directly. Refit them and measure how much they rotate between periods. If they change substantially every month, the whole framework rests on sand, and no amount of downstream sophistication will fix that.

This is the most technically demanding strategy in the mean-reversion family and the one with the widest gap between an impressive backtest and a working live system. That gap is almost entirely look-ahead, cost and crowding. Be suspicious of your own results in proportion to how good they look.