A pre-registered research record · Aahan Modi
Information-Gated Hybrid Equity Selection
Research question
What is known, and what is being tested
Observations established by records
Hypotheses not yet supported or refuted
The question
A ranking model is consistent and blind to context. A human analyst sees context and is not consistent. The study asks whether putting them together helps, or just averages two kinds of error.
Why a hybrid might succeed
Why it might fail
The experiment
Every choice that could be changed after seeing returns was fixed on and registered publicly the same day. Three analyses run side by side, and evidence is never transferred between them.
- Benchmark
- Evaluation horizon
- Prospective rounds · cutoff 4:00 PM Eastern
Locked before any outcome existed
Seven rules, each fixed on the day the protocol was frozen. None may change after a return is seen; a correction can only be appended, with its reason and checksums.
| Rule | Statement | Locked |
|---|
How each method scores a stock
Four rankings of the same twelve stocks, built from rules frozen before any outcome. Scroll to assemble them.
- 1
Quantitative model
Six factor families, standardised within sector across several hundred stocks, weighted by fixed economic priors.
- 2
Independent human score
Five criteria from minus 2 to plus 2, submitted before any model rank is visible. Confidence never changes the score.
- 3
Fixed hybrid
Both converted to percentiles, then 70 percent quant and 30 percent human, for every stock.
- 4
Information-gated hybrid
Human weight is admitted only when the Information Friction Score says the model's information is weak.
The quantitative weights are momentum 25, quality 20, value 15, growth 15, earnings information 15 and risk 10 percent. The gated hybrid uses 100/0 below an IFS of 0.40, 80/20 from 0.40 to below 0.70, and 60/40 at 0.70 or above.
Hypotheses, and what would falsify them
| ID | Prediction | Test |
|---|
Two groups, never pooled
Seven securities are really held; they were bought before the protocol existed, so they are a case study, not evidence for the model. Twelve unowned stocks form the held-out cohort that tests it. No result on this site mixes the two.
Held positions
Owned · brokerage records| Ticker | Company | Bought | Entry | Weight | To |
|---|
Held-out cohort
Never purchased · research only| Ticker | Company | Sector | Position |
|---|
The information cutoff
A forecast may only use what sits to the left of the line. Everything to the right is withheld, however obvious it looks afterwards.
Cutoff
21 Sep 2026
1 of 7 dates visible
4:00 PM ET
These are the study's real dates. Dates after are scheduled, and no outcome from them has been observed.
The prediction ledger
Now choose a cutoff yourself. You see only what was known on that day. What happened afterwards stays sealed until you ask for it, and the original record sits beside it, unchanged.
Information gate: everything dated after the cutoff is withheld, including evidence.
What happened
SealedNothing after the cutoff is shown until you choose to see it.
Every recorded prediction
| Forecast date | Ticker | Group | Quant rank | Human score | Confidence | IFS tier | Outcome |
|---|
Results
Shown only from verified observations. A result that contradicts the thesis is shown exactly the same way as one that supports it.
Where the model and the analyst disagree
Disagreements are where the study learns the most, so each one is kept, explained from both sides, and closed only by evidence.
The model saw
The analysis saw
What would settle it
Methods and limitations
What the words on this site mean, where every number comes from, and what this study cannot tell you.
Definitions
Data sources
AI contribution
Sample size and other limits
Known biases and their controls
| Bias | Control |
|---|
Protocol changes and deviations
| Version | Date | Change |
|---|
Explore the research archive
Every weekly report, every version of the paper and the data workbook, dated and kept. Nothing is replaced; new versions are added beside the old.