Key Takeaways
- +204.7% cumulative return across 107 out-of-sample trades, 2021–2025
- 1.32 Sharpe, 1.83 Sortino, –7.1% max drawdown
- 72% overall hit rate — 71.3% longs, 83.3% shorts
- Execution risk varies 120–179% between sponsors; disease area varies less than 10%
- Alpha source: divergence between quant scores and independent AI agent verdicts
- No backfilled history — all signals generated prospectively from 2020
Every clinical trial dataset used for quantitative research carries the same flaw: it reflects the registry as it exists today, not as it existed when a trade would have been placed. Endpoints get amended after interim analyses. Enrollment targets get cut after site failures. Completion dates slip. A model built on today's record is reading an edited document. AppliedXL built detection systems to track those changes in real time across 500,000+ trials. This is what a simulated trading model built on that signal produced across 107 out-of-sample trades.
+204%
Cumulative Return
107 out-of-sample trades
1.32
Sharpe Ratio
–0.01 baseline (IBB ETF)
72%
Hit Rate
55%+ is strong event-driven
107
Trades
2021–2025, walk-forward
$0.00
Net Market Beta
long/short balanced
The Data Problem
AppliedXL built detection systems that track changes across 500,000+ clinical trials — field by field, amendment by amendment, across clinical registries (ClinicalTrials.gov), scientific literature (PubMed, conference papers, biomedical ontologies), federal regulatory filings (Federal Register, 430+ agencies), and corporate disclosures (8-Ks, press releases). Custom models detect, normalize, and link events via deterministic and language-tuned pipelines. This analysis draws on 37,391 interventional trials — over one million event-level records — structured against a domain ontology built for temporal integrity.
Using the registry as it exists today is survivorship bias by another name. The document has already been revised by the outcome.
Every trade in the backtest used only information available at the moment of hypothetical entry. No future data touches any stage of the pipeline.
Two Layers of Signal
Scientific signal in clinical trials — mechanism of action, endpoint design, therapeutic indication — is actively monitored and efficiently priced. The operational layer is not. Timeline integrity, enrollment behavior, protocol amendments, and status transitions accumulate in the registry unread by any commercial provider.
Scientific Signal
Well-modeled by the market
What analysts track
- Trial design & endpoints
- Therapeutic indication
- Mechanism of action
- Scientific publications
- KOL sentiment
Operational Signal
Systematically underweighted
What no one monitors holistically
- Timeline integrity & date slippage
- Enrollment behavior vs. targets
- Status transitions and regressions
- Protocol amendments
- Dormancy and silence patterns
0% of commercial providers monitor this layer
AppliedXL's analysis across 377 trials in three indications quantified the gap. Execution risk varies 120–179% between sponsors. Disease area risk varies less than 10% across those same companies. The same companies — GSK at 0.00, Xencor at 0.45 — maintain consistent execution scores across NSCLC, Alzheimer's disease, and heart failure.
12–20×
Execution effect vs. disease effect on trial risk
Company execution capability varies 120–179% between sponsors. Disease area varies less than 10% across the same companies. The operator predicts the outcome far better than the indication.
GSK in NSCLC: risk score 0.00, zero delays.
Xencor in NSCLC: risk score 0.45, 50% of trials delayed.
Same indication. Opposite execution.
Execution capability is institutional. It persists across disease areas, pipeline stages, and market cycles. Poor operators fail everywhere. Strong operators succeed everywhere. The market prices scientific risk with reasonable efficiency. Operational risk is priced as if it doesn't exist.
Why the Signal Exists
Registry updates are made by clinical operations teams, not investor relations. FDAAA 801 mandates reporting within 30 days of any change, with penalties up to $15,000 per day for non-compliance. Companies cannot opt out. The parallel to SEC mandated disclosures is direct: Form 4 insider sales and 13F filings carry signal because they are mandatory and behavioral. CT.gov is the clinical equivalent — and receives a fraction of the analytical attention.
Mandated
FDAAA 801 requires updates within 30 days of changes. Penalties up to $15,000/day for non-compliance. Companies must report — they cannot opt out.
Timestamped
Every modification is versioned with date and time. The full behavioral history from first registration to termination or completion is preserved exactly as it occurred.
Behavioral
Registry changes reflect decisions by clinical operations teams — not investor relations. The principal-agent gap between clinical ops and IR is the signal source.
The principal-agent gap between clinical operations and investor relations is the signal source. Clinical ops updates the registry. IR manages the narrative. These are different people with different information and different incentives. Registry behavior diverges from what executives say on earnings calls, and the registry update arrives first.
What the Registry Reveals
Occurrence alone understates the signal. The same event carries different information depending on when it arrives in the trial lifecycle, how large it is relative to the original plan, whether it repeats, and what has not been reported. 1.2 million classified operational events across five years establish the empirical baseline for each pattern type.
Absence
No CT.gov updates for 6+ months
Silence is the signal.Six months without an update makes a trial 1.8× more likely to terminate. Organizational abandonment precedes formal disclosure.
Magnitude
Enrollment cut 75% vs. 20–50% vs. <20%
Size changes meaning.A 75% enrollment cut doubles failure probability vs. a minor adjustment. Same signal type, different severity.
Sequence
Delay then pull-in (whipsaw pattern)
Order reveals information flow.Delay-then-acceleration is irrational unless negative information was received between updates. 1.6× termination risk.
Distinguishing a 75% enrollment cut from a 20% one, or a delay-then-acceleration from an acceleration-then-delay, required five years of domain annotation and empirical validation against known outcomes. The classification framework is the barrier. The registry data is public. The baselines that give it meaning are not.
The Architecture
Three layers. Each feeds the next. None shares information upward until the final stage.
01
Multidimensional Prediction Engine
Five independent scoring dimensions — trial completion, timeline, primary endpoint, regulatory clearance, and clinical significance — each built from enriched registry data, SEC filings, and press releases, each answering a fundamentally different question. Not a single probability score. Five distinct failure modes, each assessed separately through domain-specific signal rules and behavioral baselines. Trial design quality carries 4× the signal weight of the next strongest feature group. The clinical significance dimension — assessing whether results will actually move markets, not just cross a p-value threshold — is, to our knowledge, original.
02
Adversarial AI Research Layer
Six proprietary AI agents run sequentially on every trial before a position is considered. A context researcher builds the factual dossier. A model narrative agent explains the prediction scores. Two adversarial researchers take opposing sides — bull and bear — searching peer-reviewed literature, FDA filings, and current press. A blind verdict agent weighs the evidence without ever seeing the model's probabilities. Only at the final stage does a reconciliation agent see everything — model scores, evidence, and blind verdicts — for the first time, adjudicating where they agree and where they diverge.
03
Readout Timing Model
Direction without timing is noise. A separate model predicts when a trial will publicly announce results — before anyone else knows — using behavioral signals embedded in registry update patterns. Timing accuracy determines instrument selection: short-dated options, intermediate options, equity, or LEAPS. Get timing wrong and even a correct directional call bleeds to theta.
Agent independence is architectural. Allowing agents to observe quant scores before forming conclusions would introduce the anchoring bias that keeps sell-side analysts clustered around consensus. The structural separation prevents it.
Feature Signal
Feature group ablation removes each cluster entirely and measures the degradation in signal accuracy. Trial design dominates across all five scoring dimensions: endpoints, controls, statistical power, and patient stratification account for roughly four times the signal weight of the next strongest group. How a trial is designed predicts outcome more reliably than what it is testing.
Feature Group Ablation
Δ AUC when group removed
Larger bar = more signal lost = feature group matters more. Negative values = removal hurts performance.
Translational biology
–0.013
Temporal trajectory
–0.006
Sponsor track record
–0.002
Chemical properties
–0.001
The Divergence Signal
Agreement between the quant scoring and the research agents generates clean signal with limited edge — the market frequently agrees too. The alpha is in divergence. Among the out-of-sample trades where the quant score was bearish and agents independently bullish, the success rate was 67 to 75 percent.
67–75%
Success rate on divergence trades — out-of-sample backtest, 2021–2025
The quant model reads patterns in structured historical data. The agents assess what is happening now — literature, filings, current registry behavior. When these perspectives conflict, the agents are reading information the historical baseline has not yet absorbed. That gap is the alpha source.
The divergence signal exists only because the agents cannot anchor to the quant score. Remove the structural separation and the signal collapses.
The Signal in Practice
Four out-of-sample calls from the prediction model — spanning endpoint, clinical significance, and regulatory dimensions. Every score was generated before the catalyst date using only data available at signal time. Returns reflect simulated trading in AppliedXL's backtest — not live positions.
NKTR · PIVOT IO-001
Nektar Therapeutics — Bempegaldesleukin
NCT04969861
P(Endpoint Met) — Out of Sample
1.7%
Simulated Short Return
+61.9%
Quadrant
Bad Drug + Bad Ops
Why Wall Street Got It Wrong
One of the most hyped immuno-oncology partnerships in history. Bristol-Myers Squibb invested $3.6B in the NKTR-214 collaboration in 2018. Multiple analysts had price targets 3–5× the eventual post-failure price.
Why the Model Got It Right
Drug score of 23.6% flagged the molecule as weak. Operations score of 1.1% was near-zero — severe trial execution red flags the hype cycle obscured. Both pillars failed independently.
NVO · SUSTAIN FORTE-2
Novo Nordisk — Semaglutide + Empagliflozin
NCT05444153
P(Endpoint Met) — Out of Sample
22%
Simulated Short Return
+22.0%
Quadrant
Good Drug + Good Ops
Why Wall Street Got It Wrong
Novo Nordisk's combination of semaglutide with empagliflozin was expected to demonstrate additive metabolic benefit. With both drugs individually proven blockbusters, analysts viewed the combination trial as low-risk.
Why the Model Got It Right
Both pillar scores were high — drug 89%, ops 94%. The 22% prediction came from cross-validation detecting that trials adding a second agent to an already-effective therapy frequently fail to show incremental benefit. A structural pattern, not a quality failure.
ACAD · COMPASS PWS
ACADIA Pharmaceuticals — Carbetocin
NCT06173531
P(Clinical Significance) — Out of Sample
13.3%
Simulated Short Return
+11.0%
Quadrant
Good Drug + Bad Ops
Why Wall Street Got It Wrong
Carbetocin targets the oxytocin receptor — a biologically plausible mechanism for hyperphagia in Prader-Willi syndrome. Analysts pointed to the unmet need and orphan economics. ACADIA's commercial track record with Nuplazid added credibility.
Why the Model Got It Right
The model scored clinical significance at just 13.3%. The same drug had already failed the identical endpoint in the prior CARE-PWS trial — a rescue attempt on a failed asset with no mechanistic differentiation. COMPASS PWS showed no separation from placebo on primary or any secondary endpoint.
UTHR · TETON-2
United Therapeutics — Treprostinil (Tyvaso)
NCT05255991
P(Regulatory) — Out of Sample
96.2%
Simulated Long Return
+33.0%
Quadrant
Good Drug + Good Ops
Why Wall Street Got It Wrong
IPF has been a graveyard for drug developers. The consensus view was that a prostacyclin analogue repurposed from PAH was unlikely to show meaningful FVC improvement in fibrotic lung disease. Analysts modeled low probability of success.
Why the Model Got It Right
The model scored regulatory probability at 96.2% — Tyvaso was already FDA-approved for PAH, and the regulatory pathway for a label expansion with strong Phase 3 data was well-precedented. TETON-2 showed 95.6 mL FVC improvement over placebo. UTHR surged 33% in a single session.
The Results
107 out-of-sample trades, 2021 through 2025. Walk-forward re-training with strict temporal holdout at each annual cutoff. Equal position sizing, no compounding, no transaction costs. All positions — long and short — are simulated within AppliedXL's backtesting framework. Every signal generated on data the system had not previously seen.
+204.7%
Cumulative Return
–13% baseline
1.32
Sharpe Ratio
–0.01 baseline
72.0%
Hit Rate
55%+ is strong for event-driven
1.83
Sortino Ratio
above 1.5 = strong downside adj.
–7.1%
Max Drawdown
worst peak-to-trough
2.92
Profit Factor
winners outweigh losers
+199.5%
Long P&L
101 trades / 71.3% hit
+5.2%
Short P&L
6 trades / 83.3% hit
–7.1% maximum drawdown against +204.7% cumulative return. The equity curve ascends consistently across a period that included interest rate shocks, FDA policy shifts, and multiple sector-wide biotech corrections. Simulated long positions produced +199.5% at a 71.3% hit rate across 101 trades. Six simulated short positions produced +5.2% at 83.3%.
Methodology Note
On methodology: Every feature passes a seven-pattern temporal leakage audit. Walk-forward validation means the system never scores a trial on data used to build its signals. No survivorship bias — failed, withdrawn, and terminated trials are in the signal development set. No parameter snooping — validation splits were established before signal construction began.
Temporal integrity is enforced structurally, not by convention. The backtester uses only the snapshot of data available at each historical entry point — registry fields, SEC disclosures, press releases — and applies no forward-looking information.
No backfilled history. AppliedXL began proprietary real-time data collection in 2020. Every signal in the dataset was generated prospectively at the moment the underlying source updated — not retroactively applied to historical records. The depth of the dataset is a function of collection tenure. There is no synthetic history.
All results are from simulated trading within AppliedXL's backtesting framework. No actual short positions were taken. Equal position sizing, no compounding, no transaction costs or slippage. Past simulation performance does not guarantee future results. This is not investment advice.