Phase 2 · deployable results
What survived the full verification stack: one axis that decides what is eligible to hold and what to avoid, and two ways — vendor and price-only — to decide how large the position should be. Neither forecasts returns for an individual name — that capability was tested for extensively and certified absent.
hi252 >= -0.12269 — keep
hi252 <= -0.4947 — avoid
Distance below the one-year high orders both outcomes monotonically. Everything else in this section is a way of cutting this ladder.
| Distance below one-year high | Rows | Bottom-quintile rate | Mean relative |
|---|---|---|---|
| Within 5% | 49,099 | 0.116 | −320bp |
| 5–12% below | 48,144 | 0.113 | −334bp |
| 11.5–22% below | 67,825 | 0.140 | −496bp |
| 22–33% below | 53,142 | 0.194 | −734bp |
| 33–45% below | 45,222 | 0.256 | −971bp |
| More than 45% below | 78,971 | 0.323 | −1,137bp |
The odds of landing in the worst return quintile nearly triple from the top bucket to the bottom. The first two rows are flat, which is what makes the ~12% cut below a sensible place to draw a line rather than an arbitrary one: the effect is flat inside 12% and degrades steadily outside it.
hi252 is the log distance below the
trailing 252-day high: log(P_today / max(P over
last 252 trading days)). It is always zero or negative. The threshold −0.12269 puts the price
at 88.5% of its one-year high or better — in plain terms, within about 11.5% of the
one-year high.
hi252 weekly for every candidate.It is an exclusion applied to a list you were already building — never a signal to buy on its own. Refresh weekly; the evidence horizon is 26 weeks.
The same axis read from the other end. Names more than ~39% below their one-year
high (hi252 <= -0.4947) hit the worst return quintile
at 0.3119 against a 0.1998 base — a 1.6× enrichment across 100,282 stock-weeks and
1,752 names, at time-z 9.94. This is the strongest time-axis evidence anywhere in the project.
Avoidance is the cheapest thing to act on: declining to buy costs nothing, so there is no transaction-cost hurdle to clear. Survivorship bias also works in its favour — the delisted names missing from the data sit disproportionately in this bucket, so the true enrichment is understated rather than flattered.
This screen was originally banked with a momentum leg attached
(mom12_1 >= 0.1764 & hi252 <= -0.4947).
Ablation killed it. Momentum alone selects names that are less likely to land in the worst
quintile (0.1812 against the 0.1998 base, time-z −2.56) — it is protective, not a risk marker. Adding
it lifts precision slightly, from 0.3119 to 0.3282, while costing ten times the coverage and halving
the time-clustered evidence, from 9.94 to 4.85. Use the drawdown leg alone.
| Set | Rows | Mean | Median |
|---|---|---|---|
| Kept — near the high | 53,505 | −264.1bp | +24.4bp |
| Excluded — deep drawdown | 53,506 | −751.7bp | −546.9bp |
| Gap | — | +487.6bp | +571.4bp |
The median gap exceeds the mean gap, and the kept set's median is positive. That is the signature of a distribution shift rather than a trimmed tail — the usual objection, that excluding volatile names mechanically lifts a skewed mean, does not apply here. A random exclusion of the same size is inert: +189.5bp at time-z 1.73, which is simply the momentum baseline.
| Cut | Gain vs universe | Time-z | Rows kept |
|---|---|---|---|
| Within 5% of high | +482.7bp | 3.75 | 29,039 |
| Within 11.5% (banked) | +436.6bp | 3.48 | 53,505 |
| Within 20% | +393.1bp | 3.71 | 70,577 |
| Within 30% | +336.6bp | 3.78 | 84,048 |
| Within 50% | +254.9bp | 2.86 | 96,842 |
Point-in-time universes, rebuilt from archived index holdings: +690bp (2022), +96bp (2024), +384bp (2025). Positive in every one, weakest in 2024.
The ladder holds across the whole liquidity range, which matters for capacity — you do not have to trade thinly-traded names to get the effect.
| Liquidity tercile | Best bucket | Worst bucket | Spread |
|---|---|---|---|
| Least liquid third | 0.118 | 0.333 | +0.215 |
| Middle third | 0.110 | 0.319 | +0.209 |
| Most liquid third | 0.120 | 0.309 | +0.189 |
The gradient itself is universe-invariant. On the same 2023+ window the S&P 500 ladder runs 0.143 → 0.344 (spread +0.202) against the Russell's 0.116 → 0.323 (+0.207) — the same per-name effect in both. A point-in-time 2022 membership audit confirms survivorship only flattens the S&P read slightly (+0.243 PIT versus +0.220 current). What differs is population: about 23% of Russell rows sit more than 45% below their high, against 2.5% in the S&P — the screen works everywhere, but the Russell offers roughly eight times as many names in the bucket worth avoiding.
An earlier version of this page attributed the flat S&P ladder to index quality-filtering. That comparison mixed date windows; matched windows show the invariance above.
peglt <= 0.57 & opinc_ttm <= 756.9
PEG-long-term in the cheapest ~26% of names and trailing-twelve-month operating income in the smallest ~14% — under roughly $757M. Small companies that look cheap relative to their growth expectations.
It predicts that the stock will land in the top tercile of 63-day realized volatility, measured 26 weeks later: 0.836 against a 0.310 base rate, a 2.7× lift, from fundamentals alone with the outcome measured on independent price data.
| Rule | Precision | Signals | Names | Time-z |
|---|---|---|---|---|
| peglt leg alone | 0.4246 | 18,266 | 225 | 10.57 |
| Size leg alone | 0.4946 | 10,319 | 127 | 16.99 |
| Both together | 0.8357 | 1,254 | 40 | 21.86 |
Neither leg alone reaches even half the conjunction's precision. This is a genuine interaction — unlike several other multi-leg rules in this project, where ablation showed a leg was decoration. Do not drop either one.
| Operating income cap | Precision | Signals | Names |
|---|---|---|---|
| $400M | 0.8643 | 774 | 24 |
| $757M (banked) | 0.8357 | 1,254 | 40 |
| $1,337M | 0.7289 | 2,778 | 71 |
| $3,000M | 0.6173 | 6,373 | 119 |
| $10,000M | 0.5139 | 12,933 | 183 |
predict 26wk-ahead tercile = current tercile
(vol: 126d estimator · liquidity: 63d · correlation: 126d)
Everything the volatility-character forecast does, at lower precision, from daily prices alone: a name's cross-sectional tercile of volatility, liquidity, and index correlation predicts its tercile 26 weeks out. No licensed columns, no vendor dependency.
| Attribute | Accuracy | Year-z | Survivorship floor |
|---|---|---|---|
| Liquidity (63d $-volume) | 0.827 | 84.9 | ~0.72 |
| Volatility (126d estimator) | 0.677 | 27.7 | ~0.57 |
| Index correlation (126d) | 0.580 | 24.0 | — |
The 126-day volatility estimator is a verified upgrade over 63 days (+2.0pp on the one-touch test, positive in all six test half-years); 252 days over-lags. The floor column is the worst case if every delisted name absent from the data (~21% of historical membership) predicted at chance.
Use the vendor rule when you hold the license — its conjunction reaches 0.836 on a targeted slice. Use this map for everything else: it covers every name with a year of price history, and its liquidity forecast is the strongest-evidenced number in the project.
Replay: gp_persist r2k _RUT ·
variant search: gp_persist improve r2k _RUT
The two rules compose naturally because they answer different questions. The drawdown axis decides what is eligible; the forecast decides how much.
The two ways of cutting the ladder are not additive — they are the same axis. Pick the one that matches how much of the universe you are willing to give up.
Replaying each rule against the banked datasets:
gp_enum eval ytr_test.tsv ytr_test_ids.csv \
"mom12_1 >= 0.1764 & hi252 >= -0.12269" rel26
gp_enum eval ytr_test.tsv ytr_test_ids.csv \
"hi252 <= -0.4947"
gp_enum eval ckrisk_y_vol_test.tsv ckrisk_y_vol_test_ids.csv \
"peglt <= 0.57 & opinc_ttm <= 756.9"