The Confluence engine is twelve layers. Each layer is a witness with its own bias, its own data source, and its own half-life. This piece walks through all twelve, in order, with the practical detail we owe anyone who wants to understand what they are underwriting.
01 — Insider buys
Source: SEC Form 4 filings. We are looking for cluster purchases — multiple officers and directors buying within a 30-day window, weighted by transaction size relative to the insider's prior holdings. A single CEO buying ten percent of his shareholding is worth more than three directors topping up by a thousand shares each. Half-life: 30 days.
02 — Activist filings
Source: SEC 13D and 13G filings. New positions above five percent, with intent disclosed. We weight by the activist's historical track record on similar disclosures — repeat winners count for more than first-time filers. Half-life: 90 days.
03 — Smart money
Source: SEC 13F filings, lagged by 45 days. We track overlap with the top decile of fundamental allocators by long-run information ratio. New entries from this set carry weight; trims do not. Half-life: 60 days, with a hard reset at the next quarterly filing.
04 — Catalyst flow
Source: USAspending and openFDA. New government contracts above a notionality threshold, FDA approval and rejection events, advisory committee meetings. The layer translates raw events into expected revenue contribution, then expresses that as a multiple of trailing four-quarter revenue. Half-life: 14 days.
05 — Politician trades
Source: STOCK Act feed. Reportable trades by sitting members of Congress. The signal is weighted by the member's historical hit rate and committee assignment relevance. Half-life: 30 days.
06 — Volume regime
Source: consolidated bid/ask tape. Relative volume vs trailing 60-day baseline, adjusted for sector flow and broad-market regime. We are looking for unusual institutional accumulation, not retail churn. Half-life: 7 days.
07 — Technical posture
Source: vendor OHLCV plus tick-level CVD. Composite of: price above 50-day and 200-day moving averages, RSI in the 40-70 band on the daily, VWAP-relative price on the 1-hour, and CVD direction on the 4-hour. The technical layer is the only layer where we explicitly mix multiple sub-signals — because none of the technical sub-signals is informative on its own. Half-life: 5 days.
08 — Quality
Source: Compustat fundamentals. Operating margin trend, free-cash-flow yield, and net leverage, scored against the 5-year sector median. Quality is a slow signal — it changes when the company changes — and we let it carry the longest half-life on the list. Half-life: 180 days.
09 — Valuation
Source: in-house DCF model. Current price as a discount to peer-median DCF and to the name's own 5-year average EV/EBIT. Like quality, valuation moves slowly. Half-life: 90 days.
10 — Earnings persistence
Source: Estimize and Refinitiv consensus history. Eight-quarter EPS slope plus rolling beat-rate. We want companies that have produced increasing earnings and have a habit of meeting or beating estimates. Half-life: 90 days, refreshed at each earnings print.
11 — News polarity
Source: newswire ingest, FinBERT-scored. Rolling 30-day consensus polarity and dispersion. We are looking for asymmetric polarity — broad agreement that the news is getting better — not the loudest stories. Half-life: 14 days.
12 — Post-earnings drift
Source: Compustat fundamentals plus price tape. Residual return after the earnings print, decay-corrected. The classic anomaly: when a company beats by enough, the drift continues for weeks. We measure the residual, not the headline beat. Half-life: 60 days.
Putting it together
The composite score is the time-decayed, IR-weighted sum of the twelve sub-scores, renormalised to 0-100. Names above 30 are on the high band; names above 40 are on the very high band. We hold the high band when our position book has room; we always hold the very high band.
The twelve are not negotiable. We do not turn them on and off. The math is the discipline.