Methodology

Market-Behavior Check

Factor attribution, the factor hedge and the risk forecast describe each style factor with about a year of daily returns. Every night we check whether the most recent weeks still look like that year. When they do not, your portfolio page says so, names the factors, and shows how much your own holdings lean on them.

Check 1: Each Factor's Recent Returns

For each factor, the last 21 trading days of daily returns are compared with the 252 trading days before them. The comparison is a kernel two-sample test (maximum mean discrepancy, MMD) with a Gaussian kernel whose width is set by the median distance between returns. The two windows are pooled and randomly re-split 1,000 times with a fixed seed. The p-value is the share of splits, the real one included, that look at least as far apart as the real split.

A factor is flagged when its p-value is below 0.01. A factor without the full 252 earlier days is not tested that night and is listed as not tested, never as passed.

Check 2: How Factors Move Together

Over the last 60 trading days, factor return correlations are turned into distances, sqrt(1 − |correlation|), and the shape of that set of distances is summarized with persistent homology. Loops that persist across scales mean the factors no longer move together in one stable pattern. This check flags when enough loops persist past a set threshold.

State Rules

The two checks answer different questions, so their numbers are never averaged. The reading combines their yes-or-no results:

StableNeither check flags.Factor numbers are read with their usual confidence.

ChangingOne check flags, or both flag on a single day.A note on your portfolio page names the flagged factors and your own score on each.

ChangedBoth checks flag on two days in a row.The portfolio note turns red and the attention queue adds a card for it.

The two-day rule keeps a one-day spike from turning the reading red.

What It Means for Your Portfolio

Your holdings each carry a score on every factor. The portfolio note shows your value-weighted score on each flagged factor, the share of your holdings' value that has a score, and the holdings that pull it most. A short position, or a fund built to move opposite its index, counts with the opposite sign. The further your score sits from the model's reference, the more of your day-to-day movement runs through the flagged factor, and the less the past year describes it.

See your own reading under “Why this signal?” on your portfolio page while a reading is active.

Data Sources

Daily factor returns from the Midas Edge stock factor model, estimated from US stock prices.
The nightly factor-stability job, which records every factor's test result, the windows, the cutoff and the date of the last market data it read.

Limitations

It is a reliability warning, not a forecast. It does not say markets will fall, and it is not a reason to buy or sell.
The test can say that recent returns differ from the past year, not why, and not in which direction they will go next.
Each factor is tested separately at a 1% cutoff. With fourteen or so factors tested every night, an occasional single flag is expected by chance alone, which is one reason a single flag reads only as changing.
The market data behind a reading is usually a day or two older than the night the check runs. The note shows the data's own date.

References

Gretton et al. (2012), A Kernel Two-Sample Test.
Zomorodian and Carlsson (2005), Computing Persistent Homology.
Edelsbrunner and Harer (2010), Computational Topology.