Methodology
The data, the formula,
and the backtest — open.
DEHY's value is structured access to every signal a fundamental analyst tracks — real-time, queryable, auditable. The DEHY Score is a triage layer on top of that data, not a price-prediction promise. This page is the receipt: where the data comes from, how the score is computed, and what the signal looks like under test — including the cuts that show no edge.
The DEHY Score
A conviction read on every insider trade — and what it isn't.
Every Form 4 we ingest carries a reproducible 0–100 score. It is a triage signal — a way to rank the hundreds of Form 4s that file each day by how worth-a-second-look they are. The high scorers surface first; the low scorers go below the fold. That alone earns its keep, before any claim about whether the score predicts returns.
We do not claim the score predicts price moves. It is versioned — algoVersion = 2026-06-03.v3 — so when the formula changes, the older numbers stay on record for comparison. The same inputs always produce the same output. There is no language model anywhere in the scoring path; it is a deterministic computation.
Five sub-factors, each 0–100, weighted into the headline
Historical pattern
The reporter's realized win rate at 90 days on their prior same-direction trades, marked to market against daily closes. A cold start — fewer than three matched samples — lands at a neutral prior of 50.
Magnitude
Shares transacted as a share of the insider's post-trade holdings. Trading 1% of a position scores low; trading 90% scores high.
Role weight
CEO, CFO, and Chair carry the most weight; non-executive officers next, then 10% holders, then directors.
10b5-1 plan dampener
Sales filed under a Rule 10b5-1 plan are heavily damped — they are scheduled weeks ahead, so their timing carries no information. Discretionary trades score full.
Cluster signal
Other insiders trading the same direction at the same issuer within a fourteen-day window.
Today · ranked by conviction
TICKR · CEO
Open-market buy
ABCD · CFO
Open-market buy
WXYZ · Director
Open-market buy
QRST · 10% owner
Disposition
A triage rank, not a recommendation. The decision stays the analyst's.
The Insider Conviction Index
One composite read per issuer, where 50 is neutral.
On top of per-filing scoring, every issuer carries a 0–100 composite — the ICI — that blends three independent signal streams into a single number.
−25 to +25
Insider component
A direction-weighted average of DEHY Scores for the issuer's Form 4 filings over the last 30 days. Buys read positive, sells negative.
−25 to +25
Institutional component
Net quarterly flow from 13F position changes. Increased and newly opened positions are bullish; decreased and closed positions bearish. Dampened when the sample is below three filers.
−15 to +15
Activist component
Recent Schedule 13D and 13G filings inside a 180-day decay window. A 13D carries full weight; a 13G is weighted 0.3×.
The composite
ici = clamp(0, 50 + insider + institutional + activist, 100)
Direction labels: bullish at 65 and above, bearish below 35, otherwise neutral.
The backtest, published openly
No edge in this regime. We say so on the record.
We test the ICI the way a diligent reader would want it tested: walk a grid of dates, partition the universe into deciles by ICI score, and measure the forward 21-day return of each decile. If the composite carries edge, the top decile should beat the bottom.
Sample window 2024-09-01 to 2026-05-13, monthly cadence across 21 sample dates. The universe is the top 150 CUSIPs by 13F holding count. The headline number — the spread between the top and bottom deciles at 21 days — came in at −0.95%.
That is a negative spread: in this bull-market regime, the composite did not separate future winners from losers at a 21-day horizon. We publish it as plainly as we would a positive result. Overclaiming buys nothing from the analysts and funds we build for; an audit trail does. The algorithm is versioned, and a future iteration can re-run and replace this figure — but only the measured number goes on the page.
What stays true regardless of the spread: the score is a triage layer. It tells a desk what to read first. The market's signals, structured at the source — that is the product, and it does not depend on a backtest going our way.
ICI decile test
21d horizonTop decile (ICI ≈ 80–100)
21-day mean return
Bottom decile (ICI ≈ 0–20)
21-day mean return
Spread (top − bottom)
−0.95%Gross event-study spread — no transaction costs, no slippage. It measures information content under ideal execution, not net trading P&L.
Data sources
Primary sources, structured the instant a filing posts.
SEC EDGAR
Form 4 (insider trades), 8-K (material events, including Item 5.02 executive changes), 10-K and 10-Q, Schedule 13D and 13G (activist and passive positions), Form 13F (institutional holdings), and Form 144 (proposed sales). Polled every 30 seconds via EDGAR’s Atom feed; the primary XML is fetched once per filing, parsed against the published schema, and stored.
Office of the Clerk, U.S. House
STOCK Act Periodic Transaction Reports. Trade dates, tickers, transaction types, and amount ranges are parsed from the official disclosures. Both the notification date — when the public can see it — and the underlying trade date are stored; tests anchor on notification date, because that is when the market could react.
CFTC
Commitments of Traders, weekly, the legacy futures-only series. Surfaced in the Macro Mirror.
Daily closes
A free, stable price endpoint, multiple years deep across covered issuers and the SPY benchmark. Used for forward-return measurement in the score and the backtest.
GDELT
Geopolitical event mentions, read-only, surfaced in the Macro Mirror as context — not as signal.
<60s
Median EDGAR-to-feed latency
30s
Discovery poll interval
Deterministic
No language model in the scoring path
Versioned
algoVersion 2026-06-03.v3
The boundaries
What we don't do — stated as plainly as what we do.
We don’t infer the intent of a trade
A 10b5-1 plan sale and a discretionary sale both file as transaction code S. We surface the code and our plan-attribution flag; you draw the conclusion.
We don’t adjust prices for splits or dividends
The price shown is the price reported on the filing.
We don’t call a model to compute scores
The same inputs always produce the same output. AI assistants can query the platform over MCP, but they read the scores — they do not set them.
We don’t generate buy or sell recommendations
The score measures conviction. The decision is the analyst’s.
Errors and corrections
EDGAR data occasionally contains errors that filers later fix via a Form 4/A amendment. We ingest amendments as separate filings and link them to the originals. We never silently overwrite an original filing with its amendment.
Coverage limits
Daily-close coverage bounds the realized-return horizon in the score and the backtest. Form 4 historical depth is still maturing — a clean "does the score itself predict returns?" study needs more years of historical fills before it can be tested without underpowered samples.
The receipts are open. The data is the product.
Real-time ingest, every form type, the API, and the MCP server.