Does factor similarity predict stock returns?

No. In a leak-free, survivor-free, total-return audit across 2005–2024 (14,181 tickers including 1,483 delisted), the cross-sectional information coefficient between factor-state similarity and forward returns is statistically zero across five methods — cosine, label-aware, supervised PLS, DTW, and walk-forward gradient boosting. Factor similarity does not forecast which stocks go up.

If factor similarity doesn't predict returns, what does it predict?

Risk. Using a stock's factor analogues to forecast its forward realized volatility shows a signal above the random baseline — cross-sectional IC +0.053 (t +7.6) across 237 monthly observations on a walk-forward embedding basis, corrected 2026-08-10 after removing a look-ahead PCA basis (the full-sample basis gave +0.067, t +10.3, in a like-for-like rerun). About 20% of the originally measured signal was artifact. Even corrected it sits well below the trivial baseline of the ticker's own trailing volatility, so it is a risk-screening signal, not a volatility forecast.

What is a leak-free, survivor-free backtest?

Leak-free means no future information enters a decision made at time t (point-in-time features, forward-only labels). Survivor-free means delisted/dead tickers are included, not just companies that still exist today — otherwise results are biased upward by silently dropping failures. Factor Weave's audit is both, plus total-return (dividends reinvested).

Is Factor Weave free, and how do I use it with an AI assistant?

There is a free tier (250 calls/day, no card). Factor Weave is available over a REST API and as a remote MCP server, so you can wire it into Claude, Cursor, or ChatGPT and ask things like 'what are AAPL's risk analogues?' — it returns live factor data for 14,684 tickers across equities, ETFs, indices, FX, crypto, and futures.

Research note

Does factor similarity predict returns? We ran the leak-free test.

We tested our own dataset the way a skeptic would — and we're publishing the result whether it flatters the product or not. The short version: factor similarity does not forecast stock returns. It does something else, something real, and that distinction is the whole point of this note.

The findings, up front

20-day forward returns are not forecastable from this factor data — cosine similarity, a return-projected (PLS) variant, and a gradient-boosted model all land at a cross-sectional IC of essentially zero, leak-free and survivor-free, across 237 monthly cross-sections and 26 years.
Factor similarity is risk-coherent (corrected 2026-08-10). Neighbours' forward volatility predicts the query's forward volatility at IC +0.053 (t +7.6) on a walk-forward embedding basis, well over the random baseline — though far below the trivial baseline of the ticker's own trailing vol (see §4). The figure previously published here was computed on an embedding basis that was not point-in-time; in a like-for-like rerun about 20% of the measured signal was look-ahead and 80% is real. Tracking issue #12.
So Factor Weave is a research substrate, not an alpha source. It is built to screen, group, and assemble clean leak-free datasets — not to tell you what happens next. Anyone selling you the latter on commodity technical factors is selling a backtest, not a signal.

1 · What's in the dataset

Factor Weave is a point-in-time factor library for US-listed equities and ETFs — roughly 10,800 tickers, daily, back to 2005 (about 28 million ticker-days). Every row is built to be joined to your own research, not consumed as a verdict.

28 daily factors

Trailing returns (1/5/20/60d), realized vol (20/60d), ATR%, RSI, 52-week z-score, momentum, mean-reversion, gap%, beta vs SPY, a composite score, plus cross-sectional ranks and quantiles.

32-D regime-aware embeddings

A vector per ticker-day summarising its factor state, with four similarity methods built on top — cosine, DTW, regime-filtered, and a return-projected (PLS) variant.

Leak-free forward-return labels

fwd_ret_1d / 5d / 20d and binary targets — each a strict price[t+h]/price[t]−1, the realised future. They are a backtest target, never an input.

Multi-asset regime classification

Each date carries a SPY-volatility regime (low / mid / high) and cross-asset conditioners — DXY, VIX, VVIX, VIX9D, TNX, XAU, VX-continuous closes + 20-day log-return z-scores, plus the VX term structure number (VX_continuous − VIX) / VIX. Any study can be split by single-asset regime or composite multi-asset state.

2 · The test, and why it's leak-free and survivor-free

The question: if you find the historical setups whose factor profile most resembles a ticker today, does the average forward return of those analogues predict the ticker's forward return? We tested it with discipline:

The embeddings use no forward labels. Candidate analogues are restricted to strictly past dates (35–400 trading days before each query).
The return-projected (PLS) model and the gradient-boosted model are refit walk-forward — each query date sees a model trained only on data whose labels were fully realised before it.
The universe is survivor-free: 14,181 tickers including 1,483 names that delisted between 2000 and today, sourced from FirstRateData's delisted archives. The most common critique of a quant null result — "you didn't measure the names that went away" — does not apply here.
Forward-return labels are computed as total return (price return + reinvested dividends from the FRD splits/dividends meta-file). Price-only labels are kept alongside for back-compat, but the headline numbers below use total return.
The metric is the cross-sectional Information Coefficient: the Spearman rank correlation between prediction and realised forward return, computed within each date, then averaged. This strips out the market-wide period effect that makes a naïve pooled correlation look meaningful when it isn't. The t-statistic comes from the spread of that IC across dates.
Scale: 237 monthly cross-sections, 2005–2024 for the similarity tests; 60 quarterly walk-forward cross-sections (2010–2024) for the gradient-boosted model. ~31 million (ticker, date) rows feed the joins.

As a sanity check every test includes a random-analogue baseline — picking K analogues at random. A correct measurement puts random at zero. Ours does.

3 · Result — returns are not predictable

Method	What it is	Cross-sec. IC	t-stat	Verdict
Random baseline	random analogues	+0.007	+1.0	≈ 0 ✓
Cosine similarity	nearest factor-profile analogues	−0.004	−0.6	no signal
Supervised (PLS)	return-projected similarity	+0.010	+1.6	no signal
Gradient boosting	nonlinear, 26 factors, walk-forward	+0.004	+0.2	no signal

Three different methods — a linear similarity, a return-supervised projection, and a nonlinear model with the full factor set — and the same answer each time: a cross-sectional IC indistinguishable from zero, and from random. The return-projected variant's mildly rising outcome buckets (a +0.70% top-vs-bottom spread) do not survive the t-test; on 237 dates it is noise.

This is not a surprise, and that's important. These are commodity technical factors on liquid US equities — the most heavily arbitraged corner of global markets. The academic and practitioner consensus is that such signals are competed flat. Our data agrees. A vendor who tells you otherwise is showing you a backtest with a leak in it.

4 · Result — but similarity is risk-coherent

Returns are a coin flip. Volatility is not — it clusters and persists. So we asked a second, fairer question: do tickers the engine calls "similar" share a forward risk profile? We measured whether cosine analogues' forward 20-day realized volatility predicts the query's, same design.

Correction (2026-08-10): about a fifth of this result was a look-ahead artifact. The rest survives.

The neighbour ranking below runs on our 32-D factor-state embeddings, and that embedding basis is currently fitted over the full sample rather than as-of each date. The inputs (trailing 64-day z-normalised return windows) are clean and the neighbour edges are strictly walk-forward, but the projection was not point-in-time — a vector dated T was expressed in a basis estimated using data from after T.

We rebuilt the embeddings with an expanding-window basis (for each year, PCA fitted only on strictly prior data) and re-ran this exact probe — same 237 query dates, same 28,440 queries, same seed. Result:

Full-sample basis (same rerun, superseded)	+0.067	t +10.3
Walk-forward basis (corrected)	+0.053	t +7.6

So 20% of the measured signal was look-ahead and 80% is real. The corrected figure is still ~21× the random baseline and still statistically significant, and it does not change the conclusion that matters: cosine similarity remains roughly 15× weaker than simply using the ticker's own trailing volatility (see the table below). This is a screening and grouping signal, not a volatility forecast. The return results in §3 are unaffected in direction — a look-ahead basis biases towards finding signal, and those tests found none, which makes the null more conservative rather than less. The production pipeline fix is tracked in issue #12; until it ships, the served embeddings and neighbour rankings carry this same caveat in their API responses.

Predictor of forward 20-day volatility	Cross-sec. IC	t-stat
Random analogues' forward vol	+0.002	+0.3
Cosine analogues' forward vol (walk-forward basis, corrected 2026-08-10)	+0.053	+7.6
Cosine analogues' forward vol (full-sample basis — superseded, contained look-ahead)	+0.067	+10.3
The ticker's own trailing vol	+0.826	+173.9

Cosine analogues carry real, significant forward-risk information — roughly twenty-one times the random baseline, t +7.6. The embeddings do organise the universe by risk regime: "similar" means "similar risk profile." Note this is the corrected figure: the figure previously published here was computed on an embedding basis fitted over the full sample (+0.067, t +10.3 in the like-for-like rerun), and re-running the identical probe on a walk-forward basis removed about a fifth of the measured signal. Four fifths of it survived.

We'll be just as straight about the limit. A ticker's own trailing volatility predicts its forward volatility far better (IC +0.83). So this is not a volatility-forecasting product — if you want one stock's vol, use its own history. What the result establishes is subtler and more useful: the similarity engine is coherent. It is not noise. When it groups setups, it is grouping them by a property that genuinely carries forward — which is exactly what makes it a sound tool for screening, peer sets, and regime-aware research, and exactly why it is not a return oracle.

5 · What to use it for

A good fit

Screening — find tickers in a given factor or risk state
Peer sets & substitutes — "what else looks like this?"
Regime-aware research — split any study by market state
Assembling clean, leak-free backtest datasets to test your own signal
Conversational research via the MCP server

Not what it does

Predict which stocks go up — the data does not contain that
Replace your alpha model — it is the substrate, not the signal
Forecast a single stock's volatility — its own history wins

Reproduce it

The probes behind every number here are in the repository under scripts/diagnostics/ — signal_probe_extended.py (similarity vs returns), signal_probe_gbm.py (the gradient-boosted model), and signal_probe_vol.py (the risk-coherence test). Leak-free by construction; run them yourself.

Honest data, honestly described

A free account gives you all 28 factors, 252 days of point-in-time history, four similarity methods, forward-return labels, the REST API and the MCP server.

Create a free account

Or explore the data first — no signup needed.