ODIN Scoring Engine
ODIN is a machine learning model that predicts FDA approval probability for biotech catalysts. Built on 2,210 historical FDA events, ODIN achieves 0.9193 AUC on walk-forward backtesting across 5 yearly folds (0.89–0.94 range). It identifies the most likely winners before they happen.
Instead of reading regulatory documents manually, ODIN automatically scores each PDUFA event across 45 features spanning 8 signal categories. Higher ODIN scores correlate with positive stock performance post-approval and larger expected moves.
Methodology & Accuracy
ODIN uses gradient-boosted decision trees (LightGBM) to predict binary FDA outcomes (approval / rejection). The model was trained on 2,210 historical FDA decisions from 2000–2025, covering NDA, BLA, and sNDA approvals across all therapeutic areas. Walk-forward testing (backtesting without look-ahead bias) validates 0.9193 AUC, confirming the model's real-world predictive power.
8 Signal Categories
| Signal Category | Description | Weight Impact |
|---|---|---|
| Regulatory Designations | Breakthrough Therapy, Fast Track, Priority Review, Accelerated Approval | High |
| Manufacturing/CMC Risk | Drug substance, drug product, analytical methods, quality assurance issues | Medium |
| Therapeutic Area History | Historical approval rates in oncology, rare disease, immunology, etc. | High |
| Sponsor Track Record | Company's historical approval rate, CRL frequency, regulatory responsiveness | Very High |
| Clinical Trial Design | Efficacy endpoints, safety profile, comparator arms, patient population | Very High |
| Prior CRL History | Completeness Response Letter issues, deficiencies, safety flags | High |
| FDA Era Effects | CDER commissioner tenure, regulatory environment, macro policy changes | Low–Medium |
| Advisory Committee Signals | AdCom vote outcome, questions asked, dissenting votes, sentiment | Medium |
ODIN Tier Definitions
TIER_1
High conviction approval. Strong trial data, solid sponsor track record, favorable regulatory path. Expected positive stock reaction.
TIER_2
Moderate–high conviction. Good data but some concerns (e.g., safety signals, competitive landscape). Likely approval.
TIER_3
Moderate conviction. Mixed data, unclear regulatory stance, or higher risk profile. Approval is possible but not favored.
TIER_4
Low conviction. Significant concerns (weak efficacy, manufacturing risk, sponsor track record). Approval unlikely.
How ODIN Works
- Data Collection: ODIN ingests regulatory filings, clinical trial results, FDA meeting minutes, advisory committee votes, and sponsor company data.
- Feature Engineering: 45 features are derived from the 8 signal categories. Each feature is selected and weighted by the LightGBM ensemble based on historical predictive power.
- LightGBM Ensemble: The trained gradient-boosted model outputs a calibrated probability score (0–100%) for each PDUFA event. Scores reflect the likelihood of FDA approval.
- Tier Classification: Scores are binned into TIER_1 (85%+), TIER_2 (65–85%), TIER_3 (40–65%), TIER_4 (<40%).
- Daily Updates: ODIN rescores all pending events nightly as new clinical data, regulatory designations, and company news arrive.
Frequently Asked Questions
How much does a higher ODIN score improve stock returns?
Historical backtesting shows TIER_1 events (85%+ ODIN scores) average +8–12% positive returns post-approval. TIER_2 averages +3–7%. TIER_3 and below show mixed results. Higher ODIN confidence correlates with larger expected moves.
Can ODIN predict rejections as well as approvals?
Yes. ODIN's 0.9193 walk-forward AUC measures performance across both approval and rejection predictions. Low TIER_4 scores (below 40%) identify high-risk rejections. When ODIN flags TIER_4, rejection probability exceeds 60%, useful for shorting strategies.
How often does ODIN get it wrong?
On validation data, ODIN has ~92% AUC (0.9193 walk-forward). This means ~12% of events fall into the 'wrong' tier. Walk-forward testing confirms 0.9193 AUC across 5 yearly folds (2021-2025). Errors tend to come from TIER_2 and TIER_3 events near the boundary.