Empirical evidence that ZenHodl's pre-game model carries information beyond luck. When our model identifies value (CLV-positive entry, defined precisely below), trades win 89.9%. When our value call is wrong (CLV-negative entry), trades win 11.2%. The 78.8-percentage-point gap is a direct test of whether the model carries information beyond luck, and we publish both halves.
The full verification script (verify_clv_gap.py) is published. It reads trades.jsonl and reproduces every number on this page.
| Bucket | n | Win rate | Mean CLV |
|---|---|---|---|
| CLV+ (close − entry > +0.5c) | 457 | 89.9% | +28.3c |
| CLV= (|close − entry| ≤ 0.5c) | 5 | 60.0% | +0.2c |
| CLV− (close − entry < −0.5c) | 493 | 11.2% | −31.1c |
Sample composition: 1,642 trades on the public ledger have a settled outcome. 955 of those (~58.2%) had a recorded closing price; the gap is computed on the 950 with a non-trivial price move (excluding 5 trades with |close − entry| ≤ 0.5c). The Z-test value above tests the hypothesis that the two win rates are equal on the aggregate dataset; it is rejected at any conventional level.
Every trade has an entry price (the market price when we fired) and a closing price (the market's final price right before the event resolved). CLV is the difference, measured in cents on a 0–100 implied-probability scale. We bucket each measured trade into one of three groups using a ±0.5c threshold:
CLV is a widely-used skill metric in sports analytics because it is judged by the market itself, not by who won the game. A bettor who consistently beats the close is identifying mispriced bets — even when those specific bets lose. A bettor who consistently loses to the close is paying retail prices regardless of how often the games go their way.
A common objection to a single headline number: "what if one sport is carrying the average?" The breakdown below shows the eight sports with at least 30 trades in the gap (CLV+ ∪ CLV− ≥ 30). Sports below this cutoff and sports with thinner pipeline coverage are surfaced separately in the selection-bias section below.
| Sport | CLV+ n | CLV+ WR | CLV− n | CLV− WR | Gap | Ratio |
|---|---|---|---|---|---|---|
| MLB | 103 | 98.1% | 92 | 12.0% | +86 pp | 8.2× |
| WTA | 59 | 91.5% | 80 | 8.8% | +83 pp | 10.5× |
| CS2 | 87 | 86.2% | 114 | 10.5% | +76 pp | 8.2× |
| LoL | 47 | 83.0% | 45 | 6.7% | +76 pp | 12.4× |
| ATP | 62 | 88.7% | 76 | 14.5% | +74 pp | 6.1× |
| NBA small-sample | 16 | 93.8% | 19 | 0.0% | +94 pp | n/a |
| NHL | 56 | 91.1% | 41 | 22.0% | +69 pp | 4.1× |
| Soccer | 15 | 73.3% | 20 | 10.0% | +63 pp | 7.3× |
Win-rate ratios across the seven sports with n_CLV− ≥ 20 range from 4.1× (NHL: 91% / 22%) to 12.4× (LoL: 83% / 7%). NBA's "n/a" ratio reflects 0 wins on n=19 CLV− trades — a small-sample artifact rather than a defined ratio. The gap is positive in every individual sport at this sample size; no single sport is carrying the headline 78.8-pp average. Per-sport win rates on samples this size carry meaningful uncertainty; the formal Z-test is computed on the aggregate dataset only, not on each sport separately.
If the 78.8-pp gap were an artifact of where we set the CLV± threshold, the gap would shrink as we move the boundary further from zero. It does the opposite — trades that beat the close by larger margins win more reliably; trades that miss the close by larger margins lose more reliably. The trend is consistent rather than perfectly monotonic at the per-bucket win-rate level (89.9% → 89.9% → 90.0% → 90.9% on the CLV+ side as the threshold tightens), but the gap grows from +78.2pp at ±0c to +82.2pp at ±5c.
| Threshold | CLV+ n | CLV+ WR | CLV− n | CLV− WR | Gap |
|---|---|---|---|---|---|
| ±0.0c | 460 | 89.6% | 494 | 11.3% | +78.2 pp |
| ±0.5c (default) | 457 | 89.9% | 493 | 11.2% | +78.8 pp |
| ±1.0c | 456 | 89.9% | 490 | 10.6% | +79.3 pp |
| ±2.0c | 451 | 90.0% | 487 | 10.5% | +79.5 pp |
| ±5.0c | 429 | 90.9% | 457 | 8.8% | +82.2 pp |
A pattern consistent with real signal rather than threshold-induced bucketing: large CLV magnitudes predict outcomes more decisively than small ones.
Of 1,642 settled trades on the public ledger, 955 (~58.2%) had a recorded closing price at settlement; the remaining 687 did not. The headline gap is computed on the measured subset only. Two questions follow:
1. Do measured and unmeasured trades win at different rates overall?
No detectable aggregate difference. Measured: 49.1% WR (n=955). Unmeasured: 48.3% WR (n=687). Two-proportion Z-test on the WR delta yields z = 0.31, p = 0.754 — well above any conventional significance threshold. The missingness mechanism does not preferentially keep winners or losers in the measured subset at the aggregate level.
2. Is missingness sport-correlated?
Yes, substantially. Coverage varies by sport. We define a sport as well-covered when ≥75% of its settled trades have a recorded closing price; partially covered when 25–75%; essentially absent when below 25%.
| Sport | Measured | Unmeasured | Coverage | Status |
|---|---|---|---|---|
| SOCCER | 35 | 0 | 100.0% | well-covered |
| UNKNOWN | 5 | 0 | 100.0% | well-covered |
| ATP | 139 | 1 | 99.3% | well-covered |
| WTA | 140 | 4 | 97.2% | well-covered |
| NHL | 97 | 8 | 92.4% | well-covered |
| LoL | 93 | 10 | 90.3% | well-covered |
| MLB | 196 | 60 | 76.6% | well-covered |
| CS2 | 201 | 192 | 51.1% | partially covered |
| TENNIS (uncat.) | 5 | 6 | 45.5% | partially covered |
| NBA | 36 | 95 | 27.5% | partially covered |
| NCAAWB | 4 | 80 | 4.8% | essentially absent |
| NCAAMB | 4 | 231 | 1.7% | essentially absent |
| TOTAL | 955 | 687 | 58.2% |
Seven sports are well-covered (≥75% coverage), three are partially covered (25–75%), two are essentially absent (<25%). NCAAMB and NCAAWB fall out of the headline analysis because the closing-line capture pipeline came online after most of our NCAA trading window had already settled. The headline 78.8-pp gap is therefore evidence of skill on the measured subset; we cannot make the same claim for NCAA from this dataset.
The full trade dataset and the verification script are published. The headline numbers are reproducible in seconds; the per-sport table, the formal Z-tests, the threshold-robustness table, and the selection-bias analysis above all come from the same script.
curl -s https://zenhodl.net/api/trades.jsonl > trades.jsonl
python3 -c '
import json
plus = minus = wp = wm = 0
for line in open("trades.jsonl"):
r = json.loads(line)
if r.get("won") is None: continue
if r.get("closing_price_c") is None or r.get("entry_price_c") is None:
continue
diff = r["closing_price_c"] - r["entry_price_c"]
won = bool(r["won"])
if diff > 0.5: plus += 1; wp += int(won)
if diff < -0.5: minus += 1; wm += int(won)
print(f"CLV+: {wp}/{plus} = {wp/plus*100:.1f}%")
print(f"CLV-: {wm}/{minus} = {wm/minus*100:.1f}%")
print(f"Gap : {(wp/plus - wm/minus)*100:.1f} pp")
'curl -O https://zenhodl.net/api/verify_clv_gap.py
curl -O https://zenhodl.net/api/trades.jsonl
python3 verify_clv_gap.py trades.jsonlThe dataset is updated continuously, so the exact numbers you see when you run the script will drift slightly from this page's snapshot. The page is dated and we re-publish a fresh snapshot roughly monthly. If your run produces materially different conclusions, please tell us.
What it proves
What it does not prove
ZenHodl. "The 78-Point Gap: Empirical Evidence of Forecasting Skill."
Snapshot 2026-05-08. Accessed 2026-05-08.
https://zenhodl.net/clv-evidencePage content licensed CC BY 4.0. Underlying dataset is public.