The 2026 RIDGE Back-test
169 historical Amazon FBA niches · 2022-2023 entry cohort · observed through 2026 · published April 2026
Headline Results
(2.04× base-rate lift)
What We Measured
We sampled 169 Amazon FBA niches where market entry occurred in the 2022-2023 window — a period far enough in the past that outcomes are now observable. For each niche we reconstructed the data RIDGE would have seen at time-of-entry, ran a verdict through the current engine, and compared that verdict to the observed outcome in 2026 (still-selling / died / commercially viable / not).
The test is deliberately brutal: the model has never seen these niches, the ground truth is observable reality (not a model prediction), and the time gap eliminates leakage. This is the standard back-test protocol used in quantitative finance, adapted to FBA.
What the Numbers Mean
- 96.2% NO-GO precision — of the niches RIDGE recommended avoiding, 96.2% turned out to actually be non-viable (n=159 NO-GO verdicts, 2,000-sample bootstrap). When RIDGE says "don't enter," you should listen.
- 41% HIDDEN GEM precision — the HIDDEN GEM signal (DIY, accessory-anchor, craft/décor categories) correctly identified viable opportunities at 2.04× the base rate. False positives exist, but the lift is real and meaningful.
- 97.8% GO precision — overall GO/NO-GO calls were correct 97.8% of the time on the 169-niche set.
- 2,710 ground-truth labels — the calibrated machine-learning verdict head was trained against an independently-held cohort and ranks borderline niches; the deterministic verdict rule remains the binary gate.
Where RIDGE Was Wrong
RIDGE is not perfect. The two places it under-performs:
- Emerging-category novelty — niches with no historical cohort are harder to score. Where the underlying category did not exist in 2022, the model leans conservative.
- Black-swan macro events — supply-chain shocks and policy changes are under-weighted if they occurred after the training window. We refresh the calibration quarterly to compensate.
These caveats are disclosed in every RIDGE report. No forecast is a guarantee; it is a probabilistic recommendation grounded in triangulated evidence.
Why No Competitor Publishes This
Publishing back-tested accuracy is expensive and commercially risky. If the number were mediocre, it would drive customers elsewhere. The typical FBA-SaaS playbook is to market on testimonials, screenshots, and case studies — all low-falsifiability marketing. RIDGE takes the opposite bet: the methodology is strong enough to publish, and customers who care about rigor will reward transparency. The numbers above are updated whenever new cohorts become observable.
Related research
- Methodology & Ablation — what each model version contributed and which experiments were rejected.
- Calibration Audit — calibrated machine-learning verdict head, audited against 2,710 ground-truth labels with bootstrap 95% CI on every metric.
- Public Leaderboard — open submission benchmark with nested temporal CV.