Verdict: Moonshot Odds Model (M-001) is rated “FIT FOR LIMITED USE — RELIABLE WITH CONDITIONS” in its validation review dated 1 Sep 2026.
Key result: Gate AUC 12m: 0.824.
Key result: prec@20 (floored): 20.0% vs 4.2% base.
How it was tested: training data 2009-01→2025-08, test data 2020-01→2025-08; out-of-sample design (how it was checked on data it never saw while being built): walk-forward yearly.
Horizon: 12m (+6m companion). Coverage: 22 markets (US + 21): listings ≥ USD 2M/day (point-in-time FX), ≥ 1y listed, US also ≥ $1; ~12,142 names scored; training panel US 2009→ (deep) + other markets 2016→ (survivor-only outside the US).
Forward record: 4 published lists are being tracked against SPY; none has reached its 252-session horizon yet (the first matures on 2 Sep 2027).
Stage: Production. Version v5; the hub record is as of 6 Nov 2026.
The verdict
FIT FOR LIMITED USE — RELIABLE WITH CONDITIONS: the monthly rank-ordering can be relied on as a candidate generator for a small, loss-tolerant, regime-gated sleeve; the levels and the probabilities cannot.
Development summary: Both v3 gates PASS: the model multiplies a 1.3-3.8% base rate 4-7× at the top of its list, but it is a rank-ordered candidate generator consumed inside the regime gate and the $10M/day liquidity floor — never as a probability.
What the review allows
Monthly rank-ordering as a candidate generator for a small, loss-tolerant, regime-gated sleeve (top-20/30 baskets).
Compressing ~3,000 eligible names to a reviewable shortlist for the momentum/speculative sleeve.
Ordering names inside the production screen: top decile ∩ regime gate ∩ ≥$10M/day liquidity ∩ fundamental overlay.
What the review rules out
Absolute backtest levels — the panel carries zero delisted names, so every level is an upper bound.
Calibrated tail probabilities — they plateau at the top; read the rank, not the number.
Performance in speculative-bust regimes — ranking was near-random 2020-22 in walk-forward stress.
Realized economics at size — v4 cost model (0.4%/side + ADV impact) shows the constraint is the $10M/day floor, not cost; capacity curve in.
Unconditional production use while N-01 (residual survivorship) stands open — every absolute figure is an upper bound by roughly 15–25% relative (v4 survivorship bound).
What it predicts
Item
Detail
Question it answers
Which listed stocks in 22 markets (US + 21) return ≥+150% over 12 months, when the unaided base rate is 1.3-3.8%?
Method
RandomForest on rank inputs
How it works
RandomForest on within-cohort rank inputs; monthly cohorts; 2009→ deep panel; walk-forward accepted
Rank-ordered monthly shortlist; rank bin -> realised lift over the cohort base rate (never a level probability)
Horizon
12m (+6m companion)
Coverage
22 markets (US + 21): listings ≥ USD 2M/day (point-in-time FX), ≥ 1y listed, US also ≥ $1; ~12,142 names scored; training panel US 2009→ (deep) + other markets 2016→ (survivor-only outside the US)
Markets
Global
Model type
Model · ML/Stat
Stage
Production
Version
v5, first recorded 1 Sep 2026
Validation evidence
Training data: 2009-01→2025-08. Test data: 2020-01→2025-08. Out-of-sample design: walk-forward yearly.
Headline results from the model's development record, as the hub publishes them.
Measure
Result
Compared with
Note
v3 rolling gate AUC (12m)
0.826
0.60 pre-registered gate
2024-09→2025-08, one honest out-of-time read before the refit
precision@20 / month (12m)
17.9%
4.9% base rate
falls to 13.7% once the production $10M/day liquidity floor is applied
decile-10 lift
3.3×
2.0× pre-registered gate
v2 OOT AUC 2024-25 (independently validated)
0.845
0.839 elastic-net logit benchmark
6-month companion, precision@20
26.2%
13.0% base rate
The hub's standard metrics for this model class (only those the record measured).
Metric
Sample / window
Compared with
Value
Range (CI)
AUC (headline)
WF 2020–25 · ≥$10M/day
n/a
0.7981
n/a
Precision at k
WF 2020–25 · top 20/mo · ≥$10M/day
2.43
17.94%
[14.34%, 21.99%]
Base rate
WF 2020–25 · ≥$10M/day
n/a
2.43%
n/a
Forward record: 4 published lists are being tracked against SPY; none has reached its 252-session horizon yet (the first matures on 2 Sep 2027).
What the validation review found
Is the design sound? Engineering core sound (270/270 replication, clean embargo). Conceptual identity differs from framing: a volatility-tier screen with thin within-tier ranking; regime features encode 2017-22 era paths; fundamentals act inversely to a quality screen.
Data quality. Labels/keys clean (0 duplicates, 0 mismatches). Total survivorship (0 delisted names). SEC coverage 59% with a time trend. y150 base rate swings 16× across years. Recycled tickers, 23 ETFs, 443 names never enter.
Backtest. Top-20 overlapping-cohort portfolio 2023-02→2026-07: CAGR 63.6% net vs SPY 20.5%, vol 53%, Sharpe 1.12 vs SPY 1.23, maxDD −36%. Walk-forward stress: AUC 0.555/0.545/0.496 for 2020/21/22; 2021 top-20 −47% (0/240 hits).
Monitoring. No monitoring layer existed; required: monthly PSI (0.1/0.25 bands), realized hit-rate loop, champion-challenger, quarterly retrain with parameterized cutoff.
Limitations
Survivorship: zero delisted names in the panel — all absolute returns are upper bounds.
Skill is regime-local: near-random ranking 2020-22 in walk-forward stress; 2021 top-20 basket −47%.
Lottery skew: the best 3 picks per cohort carry the edge; 43% of picks lose money.
Liquidity: 54% of unconstrained picks trade <$10M/day; precision@20 falls to 13.7% under the production floor.
The model is first-order a volatility ranker (59% of SHAP); AUC mostly measures vol-tier separation.
v4: level probabilities are NOT achievable — next year's base rate (0.5%-7.4%) is driven by shocks no regime state forecasts (all five calibrators fail the top-decile test on honest walk-forward scores). The model is lift-valid: its top 15% earns 3-4x the cohort base rate in every out-of-sample year; the top 5% alone ranged 0.1x-6x by year.
Questions
Is the Moonshot Odds Model validated?
Its validation review dated 1 Sep 2026 rated it FIT FOR LIMITED USE — RELIABLE WITH CONDITIONS: the monthly rank-ordering can be relied on as a candidate generator for a small, loss-tolerant, regime-gated sleeve; the levels and the probabilities cannot.
What does the Moonshot Odds Model predict?
It answers: Which listed stocks in 22 markets (US + 21) return ≥+150% over 12 months, when the unaided base rate is 1.3-3.8%? Output: Rank-ordered monthly shortlist; rank bin -> realised lift over the cohort base rate (never a level probability). Horizon: 12m (+6m companion). Coverage: 22 markets (US + 21): listings ≥ USD 2M/day (point-in-time FX), ≥ 1y listed, US also ≥ $1; ~12,142 names scored; training panel US 2009→ (deep) + other markets 2016→ (survivor-only outside the US).
How was the Moonshot Odds Model tested?
Training data 2009-01→2025-08, test data 2020-01→2025-08, out-of-sample design: walk-forward yearly. Gate AUC 12m: 0.824; prec@20 (floored): 20.0% vs 4.2% base; Walk-forward min-year AUC: 0.7965. Forward record: 4 published lists are being tracked against SPY; none has reached its 252-session horizon yet (the first matures on 2 Sep 2027).
Out-of-sample: results measured on data the model never saw while it was built or tuned.
Walk-forward: the model is refit on data up to a date and tested on the period after it, then the window rolls forward, so every test is out of sample.
AUC: area under the ROC curve: how well a score ranks cases that happened above cases that did not; 0.5 is a coin flip, 1.0 is perfect.
Base rate: how often the outcome happens with no model at all.
Precision@k: the share of the top k names on a list that met the target.
Confidence interval: the range that likely contains the true value (95% unless stated); [a, b] on this page is that range.
SPY: the SPDR S&P 500 ETF, the US market benchmark used here.
ATR: average true range: a stock's typical daily price range, often expressed as a percentage of its price.
Point-in-time (PIT): using only data that was available on each date, so a backtest cannot see the future.
Survivorship: a sample that leaves out companies that later delisted, which flatters historical results.
PSI: population stability index: how far today's inputs have drifted from those the model was built on.
Calibration: whether predicted probabilities match how often things actually happen.