Research
What Persists
Ranking thirty stocks by how well they trended tells you nothing about how they will trend next. Ranking them by how far they move tells you almost everything. One number is twenty times more durable than the other.
“Which stocks trend well enough to be worth a three-day swing?”
It is the most natural question to ask of price history, and the screen is easy to build. The hard part is the second step almost nobody runs: check whether the screen still works on data it has never seen. This one does not — rank correlation between periods is +0.04. But the same exercise on volatility gives +0.83, and that turns out to be the useful half.
See the test →The screen, over the full five years
| symbol | avg 3-day | % up | avg daily range | 3-day momentum | q |
|---|---|---|---|---|---|
| MU | +0.75% | 52.9% | 3.75% | +0.04% | 0.994 |
| PLTR | +0.74% | 54.5% | 4.80% | −0.53% | 0.994 |
| NVDA | +0.73% | 56.7% | 3.71% | −0.25% | 0.994 |
| AMD | +0.58% | 54.0% | 4.10% | +0.02% | 0.994 |
| COIN | +0.37% | 47.7% | 6.37% | −0.09% | 0.994 |
| TSLA | +0.33% | 49.9% | 4.34% | +0.01% | 0.994 |
| … | |||||
| PFE | −0.04% | 47.4% | 2.06% | +0.09% | 0.994 |
| DIS | −0.06% | 48.7% | 2.05% | +0.08% | 0.994 |
Two things are already visible. First, zero of thirty-three symbols show significant three-day momentum after correcting for the number of tests — an up move does not predict continuation anywhere, and in PLTR it predicts the opposite. So “trending” in the sense of persistence does not exist in this universe.
Second, that left-hand column is not a measure of trending at all. It is a record of which stocks went up between 2021 and 2026. The question is whether that record carries forward.
The half almost nobody runs
Build the ranking on Jun 2021 – May 2024. Then look at Jun 2024 – Aug 2026, which the ranking has never seen.
rho = +0.042, p = 0.825. Knowing which stocks trended best over three years tells you essentially nothing about the next two.
| group, chosen on training data only | test-period avg 3-day |
|---|---|
| top 5 (NVDA, COIN, AMD, COST, XOM) | +0.35% |
| all 30 symbols | +0.36% |
| bottom 5 (INTC, DIS, SOFI, PFE, GM) | +0.42% |
Picking the five best trenders returned +0.35% against a universe average of +0.36%. Picking the five worst returned +0.42%. Their individual ranks in the test period were 7th, 23rd, 4th, 26th and 16th of thirty — scattered exactly as chance would scatter them.
This is the trap that consumes screening strategies. Any ranking of thirty stocks
produces winners; a ranking is only worth acting on if it carries information forward,
and the number that decides this is rho, not the size of the gap at the top
of the table.
Something does persist, just not the part you wanted
| ranking by | rho between periods | p | |
|---|---|---|---|
| 3-day return | +0.042 | 0.825 | no signal |
| daily range | +0.831 | 1.3 × 10⁻⁸ | strongly persistent |
Volatility ranking is twenty times more durable than return ranking. The five most volatile names in the training period ranked 1st, 2nd, 3rd, 6th and 7th in the test period. COIN stays wild. SLV stays quiet. This is one of the most reliable regularities in markets, and it is the opposite of the one people screen for.
So screening works. It tells you how far a stock will move and never which way.
Why a fixed stop fails
| symbol | avg daily range | a 2% stop is… |
|---|---|---|
| COIN | 5.51% | inside one ordinary day |
| PLTR | 4.66% | inside one ordinary day |
| TSLA | 4.26% | inside one ordinary day |
| NVDA | 3.40% | inside one ordinary day |
| AAPL | 2.11% | inside one ordinary day |
Every name in the universe. A 2% stop on COIN is not risk control — it is a near-certainty of being removed by a quiet Tuesday. Tested directly on the one strategy in this project that does carry an edge, a −2%/+5% barrier pair turned +36.6 basis points into −1.9.
A stop that survives ordinary fluctuation has to scale to each stock’s own range — on the order of 1.5 to 2 times the daily range, so roughly 8% on COIN against 3.5% on AAPL. That is not a preference. It follows directly from the fact that range persists and returns do not. The same persistence is what makes volatility-scaled position sizing work: a third the size in COIN as in AAPL puts the same dollars at risk.
What would undercut this
One split, one universe
A single train/test boundary at June 2024 across thirty large caps. A different cut
date or a wider universe could move rho; it is unlikely to move it from
0.04 to anything actionable, but the number is one measurement, not a law.
Five years is one regime
The whole window is a rising market. Return persistence might behave differently through a sustained bear market, which this sample does not contain.
Three days specifically
Momentum is documented at horizons of months, not days. A null at three days says nothing about twelve-month momentum, which is a genuinely different and better-evidenced effect.
Volatility persistence is not a strategy
Knowing COIN will keep moving more than SLV is a sizing input. On its own it earns nothing.
Forward three-day returns per symbol, market-adjusted against SPY. Momentum is measured as the forward return conditional on a positive trailing three-day move, against the symbol’s own unconditional distribution, corrected with Benjamini–Hochberg across all symbols. Persistence is Spearman rank correlation between the two periods.
This is a measurement of historical data. It is not investment advice, a forecast, or a recommendation about any security.