Important: PDUFA.BIO is not the FDA. ODIN scores are ML model outputs, not financial advice. All investing carries risk. Read full disclaimer →

PDUFA.BIO
Tue Mar 3, 2026
12 this week

About ODIN — FDA Approval Prediction Engine

ODIN is a machine-learning scoring engine purpose-built for one job: quantifying FDA approval probability for upcoming PDUFA catalyst events. It exists because retail biotech investors face the same binary risk events as institutional desks, but without systematic tools to assess them.

Every PDUFA date is a coin flip for most investors. ODIN turns that coin flip into a weighted probability by analyzing 54 distinct signals across regulatory history, manufacturing quality, clinical endpoint strength, therapeutic area risk, and FDA era effects. The current v1108 engine achieves 0.88 AUC on validation and 0.91 AUC on walk-forward testing. The core principle: the RUNUP is the trade — never hold through the binary event.

How It Works

At its core, ODIN is a 54-weight logistic regression trained on 2,210 historical FDA PDUFA events (713 CRLs, 32.3% base rate) from 2015 through 2026. The current architecture (v1108) uses 33 canonical weights, 15 era/extension weights, and 6 interaction terms, optimized via random perturbation + selection on validation Brier score (0.102 val, 0.088 walk-forward).

For each upcoming catalyst, ODIN starts with a base probability derived from the overall historical approval rate, then adds or subtracts weighted signals based on the specific characteristics of that drug, company, and regulatory context. The raw logit score passes through a sigmoid function to produce a final probability between 0% and 100%.

// ODIN scoring formula
logit = base_logit + Σ(signal_weighti)
probability = 1 / (1 + e-logit)
// 54 weights: 33 canonical (v1066) + 15 extension (v1070) + 6 interactions (v1108)
// Tier assignment: T1 (≥85%), T2 (65-85%), T3 (40-65%), T4 (<40%)

Catalysts are then assigned to conviction tiers that drive position sizing and risk management. Tier 1 events represent the highest-conviction opportunities. Tier 4 events carry enough red flags that ODIN recommends no position.

54 Weights Across 8 Signal Categories (v1108)

Regulatory Designations (5)
Priority Review (+0.617)
Breakthrough Therapy (+0.166)
Orphan Drug (+0.154)
Fast Track (+0.198)
Accelerated Approval (+0.520)
Sponsor & Application (5)
Experienced Sponsor ≥5 approvals (+0.871)
Inexperienced Sponsor 0 approvals (-1.192)
sNDA/sBLA Penalty (-0.410)
Class 1 Resubmission Boost (+0.538)
Class 2 Resubmission Penalty (-0.200)
Therapeutic Area Risk (7)
TA Base Score Adjustment
High-Risk TA Penalty (-0.247)
Moderate-Risk TA Penalty (-0.178)
Low-Risk TA Boost (+0.091)
Very High TA Boost (+0.200)
Indication Pain Penalty (-0.355)
Indication Oncology Boost (+0.254)
CMC / Manufacturing (5)
Manufacturing Risk (-1.012)
Form 483 Observations (-1.315)
EMA CMC Flag (-1.736)
CMC Extension (-1.008)
Missing Pediatric PK (-1.868)
CRL & Safety (6)
Prior CRL (-3.210)
CRL Count Penalty (-0.30/CRL)
Double CRL Penalty (-0.500)
Safety Severity (-0.25/level)
Psychedelics Penalty (-1.500)
PPM Flag (-0.150)
FDA Era & Market (6)
Hoeg Era Constant (-0.100)
Accel Approval Post-2024 (-0.300)
EU Approved Boost (+0.200)
EU Approved 2026 Reduction (-0.150)
Experienced Sponsor 2026 Reduction (-0.100)
BTD Oncology Boost (+0.150)
Advanced Signals (8)
Insider Trading Signal (s23)
Hiring/Glassdoor Signal (s6)
Social Sentiment (LunarCrush)
ODIN Prior Estimate
Historical TA Approval Rate
Novice x High-Risk TA (-0.420)
Gene Therapy Penalty (-0.800)
AdCom Vote (High +1.419 / Mid -0.485 / Low -0.787)
v1108 Interaction Terms (6)
Prior CRL × Mfg Risk (-0.50)
Prior CRL × Form 483 (-0.50)
Gene Therapy × Mfg Risk (-0.80)
Inexperienced × Mfg Risk (-0.70)
Single-Arm × Surrogate (+0.40)
BTD × Single-Arm (+0.50)

How ODIN Evolved

34
v10.66 Base
34 parameters
Core logistic regression model trained via GPU on 2,200+ historical PDUFA decisions and 2,000+ phase readouts (2015-2026). Captures regulatory designations, sponsor experience, therapeutic area risk, and application type signals.
38
Round 1: CMC Forensics
38 parameters
Deep-dive into manufacturing and chemistry signals. Added Form 483 inspection flags, EMA CMC warnings, PDUFA extensions for chemistry issues, and Section 22 pediatric PK requests. Identified that CMC problems are the #1 hidden cause of CRLs for small-cap biotechs.
42
Round 2: Endpoint + Landscape
42 parameters
Forensic analysis of remaining backtest errors revealed endpoint quality and competitive dynamics as key blind spots. Added primary endpoint miss history, surrogate-only endpoints, single-arm pivotal risk, first-in-class advantage, unmet medical need, and me-too competitive penalty.
46
Round 3: Interactions + Division + Social
46 parameters
Non-linear interaction terms for compounding risk (inexperienced sponsors with manufacturing problems). FDA division-level risk adjustments for historically favorable vs. stringent review divisions. LunarCrush social sentiment integration for real-time market signal on high-conviction catalysts.
51
Round 4: Post-Mortem Driven
51 parameters
Five new signals derived from IRON/Bitopertin CRL post-mortem (Feb 13, 2026). Accelerated approval pathway risk for non-oncology filings, surrogate-clinical endpoint disconnect quantification, designation-surrogate interaction penalty, repurposed failed compound history, and early FDA action risk detection. First real-world CRL validation: ODIN surrogate_only signal exactly matched FDA rationale.
63
Round 5: Comprehensive Audit
63 parameters
Twelve new signals from retrospective scoring of 22 FDA binary events (Sept 2025 – Feb 2026). Split CRL types: CMC-only CRLs now penalized -1.0 vs -3.35 for efficacy CRLs, fixing 3 false negatives (MIST, OMER, FBIO). Added Hy's Law detection (-1.5) and FDA benefit-risk negative signal (-1.5), fixing Sanofi/tolebrutinib false positive.
54
v1108: Production Engine
54 parameters
Complete rewrite to 54-weight logistic regression optimized on 2,210 PDUFA events. Three weight layers: v1066 canonical (33 weights, immutable), v1070 extensions (15 weights for era effects, gene therapy, psychedelics), v1108 interactions (6 compound features). New tier thresholds: T1≥85%, T2≥65%, T3≥40%. Honing engine v5.1 with sign constraints and random perturbation optimization. Validation Brier 0.102, AUC 0.88. Walk-forward AUC 0.91. Gungnir stacker (XGB+Logistic) runs parallel for highest-conviction calls. VNDA Bysanti: Gungnir 89.7% TIER_1 called 4 days pre-decision.

What ODIN Is Not

ODIN is a statistical tool, not a crystal ball. It quantifies historical patterns — it does not predict the future. FDA decisions are influenced by factors no model can fully capture: advisory committee dynamics, post-submission safety signals, political considerations, and manufacturing inspections that happen behind closed doors.

ODIN does not provide financial advice. Probability scores are statistical estimates based on publicly available historical data. They are not recommendations to buy, sell, or hold any security. Past approval rates do not guarantee future results. Always consult a qualified financial advisor before making investment decisions.

Binary catalyst events can result in extreme price volatility regardless of what any model predicts. Never invest money you cannot afford to lose.

PDUFA.BIO

FDA catalyst intelligence powered by the ODIN v1108 scoring engine. Trained on 2,200+ PDUFA decisions & 2,000+ phase readouts (2015–2026).

PLATFORM

LEARN

PDUFA.BIO is the data-driven FDA PDUFA calendar and biotech catalyst calendar built for quantitative investors. Track upcoming PDUFA dates for 2026, FDA drug approval action dates, and biotech earnings dates in a unified, filterable calendar. Learn what a PDUFA date is and how the ODIN AI scoring engine generates FDA approval probability scores. Use the biotech catalyst screener to filter by ticker, therapeutic area, and ODIN tier. Explore the PDUFA runup strategy and verify ODIN's accuracy on the verified track record page. Browse our biotech glossary and catalyst research hub.

DISCLAIMER: PDUFA.BIO is not affiliated with, endorsed by, or connected to the U.S. Food & Drug Administration (FDA) or any government agency. The name "PDUFA" refers to the Prescription Drug User Fee Act and is used descriptively. This site does not provide financial, investment, legal, or medical advice. PDUFA.BIO is not a registered investment advisor, broker-dealer, or financial planner. Probability scores are generated by machine-learning models based on historical data and publicly available information. These scores are statistical estimates, not predictions or guarantees of FDA action or securities performance. Past approval rates do not guarantee future results.

RISK WARNING: Investing in biotechnology and pharmaceutical securities involves substantial risk, including the risk of total loss of capital. Binary catalyst events (such as PDUFA dates) can result in extreme price volatility. You should not invest money you cannot afford to lose. Always consult with a qualified financial advisor before making investment decisions. PDUFA.BIO, its operators, contributors, and affiliates accept no responsibility or liability for any losses arising from use of this site.

© 2026 PDUFA.BIO — All rights reserved.Market data via FMP · Social data via LunarCrush