⚠ Important limitations: This backtest has look-ahead bias — it assumes you knew investor holdings on the quarter-end date, but a 13F is not public until it is filed, which across 3,507 real filings is a median of 45 days after the quarter it describes. Exactly one of those 3,507 was public within the 3 days this tool assumes. It also has survivorship bias: the tracked investors were selected because they're already famous winners — a list picked with hindsight will look better than one picked in advance. Transaction costs, taxes, and slippage are not modelled. Past results do not predict future performance. This is an educational tool, not investment advice.
What the confidence score actually predicts
We ran the model over 155,019 stock-quarters across 46 quarterly
dates from 2015 to 2026, entering each position on the day the filing became
public rather than the day the quarter ended. That removes the bias the tool
above still carries. A placebo test — the same analysis with the answers
shuffled — reads −0.002, which is how we know the result is
not an artefact of the method.
Score
Beat the S&P over 3 months
Average excess return
Cases
0–25
39.9%
-0.66 pp
100,496
25–45
45.3%
-0.09 pp
39,662
45–65
46.4%
+0.06 pp
11,720
65–85
47.1%
+0.08 pp
3,010
85+
Never measured — fund breadth alone cannot reach this range.
No band beats the index more than half the time. That is not
a failure of the score — single-stock returns are right-skewed, so a minority
of big winners carries a cap-weighted index and the median stock loses to it.
The signal is the gap between 39.9% and 47.1%, not the level.
Three things this measurement found that we would rather it had not
The edge is concentrated before 2021.
Rank correlation with 3-month returns was +0.040 over 2015–2020 and
+0.010 over 2021–2026 — the recent half is not statistically
distinguishable from zero. We checked whether this was an artefact of a
growing universe or of market conditions; it is neither.
It works in rising markets and reverses in falling ones.
+0.035 across 37 up quarters, −0.018 across 9 down quarters. Part of what
the score measures is market exposure rather than stock selection.
It is better at what to avoid than what to buy.
Excluding the lowest-scoring decile was worth about +0.99 points a year
and worked in 10 of 12 years. Holding the highest-scoring decile was
worth half that and failed in 4 of 12 — and it has never been the
best-performing decile at any horizon.
Method, code and full results: research/04_verdict.md and
research/BUNDLE_lag50.json in the repository. Measured
2026-09-14.
Configure backtest
Fetching historical filings & stock prices…
First run for a quarter may take 30-60 seconds. Results are cached after.
Stock-by-stock returns
Model vs S&P 500
Stock-by-Stock Breakdown
#
Company
Ticker
Sector
Conviction
Buy price
Exit price
Return
⚠ Disclaimer: This backtest is purely hypothetical. No real trades were or are being made. Adrian Bogen is not a financial advisor. Past model performance does not predict future results.