Learn Lesson 3 of 6 8 minutes
Does any of this predict which fund wins?
Short answer: no. Not to any degree that survives scrutiny, and not by enough to be worth your attention. This lesson is the evidence — because you can order every list on this site by every measure we derive, and you should know what an order is worth before you use one.
Built on Every measure we hold, tested over ten years.
The result
Nothing predictedOver ten years, 646 equity funds and 33,299 fund-date observations, no signal we can compute on the day of a decision separated tomorrow's winners from tomorrow's losers once the statistics were done honestly.
The one number that did hold up is not a forecast at all: choosing the Direct plan over the Regular plan of the same fund was worth a median 1.21% points a year (middle half 0.91 to 1.58, measured on 33,028 fund-dates). That single choice is larger than every skill signal in the study put together, and unlike them it is arithmetic on a fee rather than a prediction.
docs/PREDICTIVE-STUDY.md · panel of 120 monthly formation dates, 2016-09 to 2026-08 · the cost lesson uses that one number
How it was tested
120 times — once a month from 2016-09 to 2026-08 — the study reconstructs what was knowable about every equity fund on that day: every disclosure up to that date and none after, every NAV up to that month, every factsheet already published, every manager stint that had started and not visibly ended. Funds that later closed or merged stay in — all 60 of them — so the test cannot be won by quietly dropping the failures. Each date is then collapsed to one observation, because 33,299 overlapping fund-months are not 33,299 independent facts, and the error bars are computed on the dates rather than on the rows.
The measure of success is the quintile spread: rank every fund by the signal, take the top fifth against the bottom fifth, and ask how far apart their next twelve months were. A signal that works shows a positive spread larger than its error bars. A deterministic hash of the fund id runs through the identical pipeline as a control — if a real measure cannot beat the hash, it has not found anything.
What each measure scored
Twelve-month horizon, against the median fund in the same peer group. A positive spread means the measure pointed the way we expected it to.
| Measure | Top fifth minus bottom fifth | t | p | 95% interval | Top fifth beat the median |
|---|---|---|---|---|---|
| How often it beat its category (P4, 3-yr windows) p4WinShare36 | −0.68% points a year | -0.59 | 0.557 | −2.63 to +1.81 | 52% |
| Top-ten picks that beat their peers (P3) p3Top10HitRate | −0.66% points a year | -0.47 | 0.635 | −3.42 to +2.17 | 53% |
| How long top holdings are kept (P2) p2PersistenceMonths | +0.19% points a year | 0.33 | 0.739 | −0.95 to +1.24 | 52% |
| Turnover (P1) p1TurnoverAnnual | −1.72% points a year | -2.39 | 0.017 | −3.10 to −0.28 | 45% |
| Regular-plan mark-up (P5) p5DirectRegularGapPp | −0.59% points a year | -0.55 | 0.584 | −2.72 to +1.52 | 53% |
| Size within its category (P6) p6AumPercentile | +1.18% points a year | 1.56 | 0.120 | −0.28 to +2.71 | 53% |
| Lead manager tenure (P7) p7LeadTenureMonths | −0.53% points a year | -0.56 | 0.574 | −2.34 to +1.35 | 60% |
| Expense ratio terBp | −3.52% points a year | -2.05 | 0.040 | −7.08 to −0.27 | 50% |
| Its own past one-year excess return past1yExcessPp | +0.24% points a year | 0.11 | 0.914 | −4.35 to +4.44 | 54% |
| Its own past three-year excess return past3yExcessPp | −0.44% points a year | -0.17 | 0.863 | −4.95 to +4.93 | 52% |
| A random number, as a control randomSignal | −0.19% points a year | -1.07 | 0.283 | −0.54 to +0.16 | 52% |
| All of them combined into one score combined | −0.24% points a year | -0.34 | 0.733 | −1.55 to +1.21 | 48% |
22 signals were tested; 12 of them are shown here. Correcting for having tested 22 things at once (Benjamini–Hochberg at 5%), nothing survives. Every top-quintile hit rate sits within a few points of 50% — picking from the top fifth by any of these is close to a coin toss. The two signals with small p-values point the wrong way, and neither is a claim about skill.
Read the row for a fund's own past excess return before any of the others. Trailing performance is the measure every fund platform sorts by, and its top-fifth-minus-bottom-fifth spread here is a quarter of a percentage point with an error interval ten times as wide. That is the single most useful line in the table.
Why a past-performance table finds a pattern that is not there
Three things do the damage, and the study met all three. What makes them teachable is that we can show you our own failures rather than a textbook's.
1. Overlapping windows
Roll a three-year window forward one month at a time and consecutive windows share thirty-five of their thirty-six months. A hundred of them look like a hundred observations and behave like a handful.
In our own study: At three years our own table showed four measures clearing the significance bar. Re-measured only on dates a whole horizon apart — the honest count — the strongest fell to a t of 1.41 and the next was not computable at all.
2. Testing many things
Test twenty-two measures at a 5% threshold and roughly one will clear it on noise alone. Report only the one that cleared it and you have published the noise.
In our own study: Two of ours crossed p < 0.05, both pointing the wrong way, and neither survived the correction for having tested twenty-two things at once.
3. Survivorship
A list of funds that exist today has quietly dropped the ones that merged or closed, which are not a random sample of the ones that did not.
In our own study: We kept all sixty of them. It moved every three-year spread by less than a third of a point, in both directions — a smaller bias than is usually claimed, and worth saying because we measured it rather than assumed it.
The falsification row
p = 0.043A deterministic hash of the fund's own identity — a number containing no information whatsoever — crossed the conventional significance threshold at the three years horizon against the cap-matched index blend.
Nothing was wrong with the pipeline. That is simply what happens when a three-year horizon is measured on monthly dates: ten years of data contains 2 independent three-year windows and 1 five-year one, and no statistical correction can manufacture evidence that is not in the sample. The family-wise correction refuses the hash, which is exactly why it is not optional — and why any three-year or five-year claim in this business deserves the same suspicion.
The same control at twelve months, where there are 9 independent windows, scores nothing at all.
What the best ranking in the sample actually delivered
"What percentage upside" is the wrong shape for this data, so here is the right one. The top decile of the best-performing ranking in the sample, twelve months on, against the median fund of its own peer group: a median of +0.68% points a year, a one-in-four chance of trailing the median fund by more than 6.49 points a year, and a one-in-ten chance of trailing it by more than 12.46. It beat the median fund 52% of the time.
The hash's top decile delivered a median of +0.14% points a year and beat the median fund 51% of the time. Read those two sentences together and you have the finding.
What we therefore do
Every measure on this site is published as a description of what has already happened to a named fund, with the sample stated and the counter-argument beside it. You can order any list by any of them, because being able to see who is at each end of a measure is genuinely useful. The default order is size — a fact, not a claim. And the one number we will put in front of you unprompted is the cost of the Regular plan, because it is the only one the evidence supports.
What we will not do
No "best funds". No score that claims to identify likely outperformance. No expected return, target, or percentage upside. No ordering of any list by trailing return, which is the ranking this study exists to refuse. And no combined score in any visible form — ours performed indistinguishably from a hash of the fund's own id.
What would change this
A measure we do not yet compute; a longer archive, which arrives a month at a time and gains one independent three-year window every three years; or a horizon and a peer definition under which something holds up in a test written before the answer is known rather than after. A genuinely different class of data would help more than another signal from the same disclosures — the falsification row is a standing reminder that adding the twenty-third signal to a family of twenty-two makes a spurious discovery more likely, not less.
Earlier, at the level of single stocks: the curation score's top-minus-bottom spread was roughly zero at 21, 63 and 126 trading days and −3.9 percentage points at 252 days, across 95,871 scored events and 163 cycles. Same conclusion, one level down.
The argument against this lesson
A negative result is a statement about one sample, one decade and one country's disclosures. It does not prove that nothing can ever be found — it proves that we did not find it, in data that ends in 2026-08, over a decade in which mid and small caps beat large caps by a wide margin and dragged every measure correlated with going down the cap scale along with them. Somebody with a different class of data, or with the patience to wait for more independent windows, may yet find something. What the study does establish is narrower and firmer: none of the measures we publish earned the right to be presented to you as a forecast, and a list sorted by any of them arranges the past rather than anticipating the future.
Everything above is a description of what has already happened, measured from what fund houses, AMFI and the exchanges publish. It is not advice, not a recommendation to buy, sell or hold anything, and not a forecast — neither Anveshan nor Lineage Money is a SEBI-registered investment adviser or research analyst.