We Tested Whether You Can Follow Smart Money. You Mostly Can't.

June 11, 2026 · FindataFox

An out-of-sample study of 9,240 institutional 13F filers — and why "follow the best funds" doesn't work.

TL;DR

Every hedge-fund-tracking site is built on one promise: find the funds with the best track record, copy their trades, beat the market. We had the data to test that promise directly. So we did — the honest way, out-of-sample.

The result: a fund's past skill barely predicts its future skill. Sort funds by their historical factor-adjusted alpha and the best-ranked group goes on to earn the same future return as the worst-ranked group. Of the funds that looked genuinely "skilled" in the first half of their history, only 1 in 125 still looked skilled in the second half.

This isn't a knock on any individual manager. It's the well-documented reality of public markets — and it's the opposite of what the "smart money" industry sells you.

The promise everyone sells

Dozens of hedge-fund-tracking sites and newsletters: the pitch is always some version of "these funds are the smart money — here's how to ride their coattails." It's intuitive. Institutions have armies of analysts, expensive data, and decades of experience. Surely the best of them have an edge you can borrow.

The problem is that "they have an edge" and "you can profitably copy their edge from a public filing" are two very different claims — and almost nobody tests the second one. We did.

Before any of this works: the data is a mess

The Securities and Exchange Commission has not necessarily reviewed the information in this filing and has not determined if it is accurate and complete. The reader should not assume that the information is accurate and complete.

That's not our disclaimer — it's printed on the cover page of every 13F filing. The regulator itself tells you not to assume the data is accurate or complete. There's a step the "smart money" pitch always skips: 13F filings are far dirtier than they look, and if you don't clean them, every number downstream is wrong. This isn't an edge case — it's the everyday state of the raw data every "follow the funds" product is built on. A few of the potholes we had to fix before computing a single alpha:

• The same company, filed dozens of ways. Apple alone appears under 30+ different issuer names in the raw filings — APPLE INC, Apple Computer Inc (a name it dropped in 2007), AppleComputerInc, N/A, and outright blanks. Group naively and you shatter one stock into many.
• Options masquerade as stock. A fund's put on NVIDIA is filed under NVIDIA's CUSIP, so a naive parser folds PUT NVIDIA CORP $125 EXP … into NVIDIA's ownership and "conviction." Bullish-looking positioning can literally be a bearish hedge.
• Double-counting from related filers. BlackRock Inc and BlackRock Fund Advisors both file 13Fs reporting the same underlying shares. Sum them and a stock's "institutional ownership" sails past 100% of its float — a number you'll see on plenty of sites.
• CUSIPs that don't resolve. Roughly half of 13F line items don't map to a current ticker out of the box — delisted, merged, reorganized, private placements, foreign securities, even the same CUSIP filed in different letter cases.
• Units that lie. Some values are in thousands, some in full dollars; some share counts are scaled, amendments restate prior quarters. Miss one convention and a position is off by 1,000×.

None of this is exotic. It means a lot of the "signals" you see elsewhere are partly artifacts of dirty inputs, not real institutional behavior. We rebuilt the pipeline to handle each of these before measuring anything — and even with the data properly cleaned, the answer to "can you follow it?" was still no. If the inputs aren't cleaned, the case for following smart money is weaker still.

How we measured "skill" honestly

13F filings tell you what large institutions held at the end of each quarter. From roughly a decade of those quarterly filings (~45 quarters) we reconstructed each fund's quarterly stock returns and asked a sharper question than "did they beat the market": did they generate return that isn't explained by simple style tilts?

A fund that just buys small-cap value stocks will beat the S&P in a small-cap-value year — but that's not skill, it's a factor exposure you could buy for a few basis points in an ETF. So we ran a Fama-French 4-factor regression on every fund, stripping out market, size, value, and momentum exposure. What's left — the regression intercept — is the closest thing to genuine stock-picking skill: "Skill Alpha."

The credibility test we had to pass first

Early versions of this metric were wrong, and the way they were wrong is instructive. If you price a fund's end-of-quarter holdings over the quarter that just happened, you retroactively credit them for owning the winners — a look-ahead bias. Under that flawed method, pure index funds like Vanguard and BlackRock showed +4–5%/yr of "significant skill" — which is impossible; they don't pick stocks.

We fixed it by pricing each fund's prior-quarter holdings going forward (a real buy-and-hold an investor could have replicated), plus corrections for dividends and price coverage. The tell that the fix worked: Vanguard's and BlackRock's skill alpha collapsed to ~0 and became statistically insignificant — exactly what they should be. Only once the index giants read as "no skill" did we trust the numbers for everyone else.

After de-biasing, the median fund's skill alpha is about −1.3%/yr — most institutions, net of their style tilts, modestly trail. About 989 of 9,240 funds clear statistical significance — more than chance, so some real skill exists. The question is whether you can use it.

The actual test: does skill persist?

Here's the key move. Everything above is backward-looking — a historical fit. "Smart money" is a forward claim. A backward fit is worthless to you unless past skill predicts future skill.

So we split each fund's history in two:

• Formation period (earlier half) — measure Skill Alpha here.
• Holding period (later half) — measure Skill Alpha again, independently.

The holding period is never used to pick the fund, so it's a genuine out-of-sample test: if skill is real and repeatable, funds that scored well in formation should keep scoring well in holding. (Restricted to funds with ≥16 quarters of history so each half has enough data: n ≈ 6,902 funds.)

We checked three things.

What we found

1. The correlation between past and future skill is a whisper

Spearman correlation between formation-period and holding-period skill: ρ = +0.11.

It's technically positive, and technically "significant" — but only because 6,902 is a huge sample. The effect size is tiny: past skill explains roughly 1% of the variation in future skill. As a basis for picking funds to follow, that's noise with a rounding error of signal.

2. Sorting by past skill does not separate future winners from losers

We sorted funds into quartiles by their past (formation-period) skill, then compared that ranking to what each quartile actually earned in the future — the out-of-sample holding period:

Read that again. In the past, the best quartile beat the worst by a yawning 16 percentage points — that's the ranking a "follow these funds" product is built on. Going forward, that same best-versus-worst gap is +0.02 points: the funds with the best historical skill and the funds with the worst earned the same future return. The middle quartiles aren't even monotonic. The ranking has no predictive power.

An analogy that makes it click. Rank 6,902 coin-flippers by how many heads they got in their first 100 flips. The "best" flippers and the "worst" flippers will both flip about 50% heads going forward — identical. Their past "success" was luck, not a skill that carries forward. Our funds behave the same way: the historical ranking looks meaningful, but it doesn't survive contact with the future.

3. "Proven" funds revert — the winner's curse

The strongest version of the test: take only the funds that were significantly skilled in the formation period — the ones a track-record screener would crown.

By significantly skilled we mean a fund whose formation-half Skill Alpha was both positive and statistically significant (p < 0.05): its factor-adjusted edge was above zero, and big enough relative to its quarter-to-quarter noise that it's unlikely — less than a 1-in-20 chance — to be luck. (A large alpha with a weak t-stat doesn't count; that's exactly the kind of number that's usually luck.) That's a deliberately high bar — not "beat the market once," but "showed a real, measurable, statistically-distinguishable edge across the first half of its history." Of the 6,902 funds, 121 cleared it.

In the out-of-sample holding period:

• Only 43% stayed even positive (a coin flip).
• Just 0.8% — one single fund out of 121 — stayed significantly skilled.

Past "proof" of skill almost entirely evaporated when we looked forward.

What this means

This matches 30 years of academic finance — Carhart's 1997 mutual-fund work and everything since: manager outperformance rarely persists. What looks like a hot hand is mostly the survivors of thousands of funds taking thousands of bets; with ~9,000 funds, hundreds will look "significant" by luck alone.

For you, the practical takeaways:

• "Follow the funds with the best track record" is not a strategy. Our own data, tested the fair way, says the ranking doesn't carry forward.
• A high alpha with a borderline t-stat on a few years of data is the most likely to be luck — and the most likely to be sold to you as skill.
• The famous names are mostly closet indexers in their 13F books. Vanguard, BlackRock, Geode all read as market-beta ≈ 1, skill ≈ 0. You're not missing a secret; there often isn't one.

The honest caveats (because the honesty is the point)

We hold ourselves to the standard we're criticizing others for skipping:

• 13F is a partial view — long U.S. equity only. No shorts, options, bonds, international, or cash. A macro or market-neutral fund's real edge can live entirely outside what we can see.
• Quarterly snapshots miss intra-quarter trades. We assume the filed book was held all quarter.
• We measure the fund's own skill, not a follower's. A copycat couldn't act until the 13F is public ~45 days into the next quarter — a stricter "could I have cloned this?" test would look even weaker, not stronger.
• We validated the benchmark leg against known returns, but never against a fund's actual NAV. Treat these as a rigorous proxy, not an audited track record.

None of these caveats rescue the "follow smart money" thesis — most of them make it harder, not easier.

Methodology and limitations in full: see our Alpha Methodology reference. This is research and education, not investment advice. Past performance — as this entire piece argues — does not predict future results.