# ADR-0012: Confluence Signal Engine (M22)

Date: 2026-08-10  
Status: Accepted

## Context

Investor Flow evaluates entry/exit quality for a symbol by combining multiple
evidence axes (technical, institutional, macro, seasonal, flows, sentiment).
The existing codebase had per-indicator helpers (RSI, MACD, moving averages,
volume-by-price, rotation, seasonality) but no unified layer that combined
them into a single, per-symbol, per-date assessment. Users saw individual
indicators but not the synthesized picture.

The product needs an evidence-based confluence layer that:
- Aggregates multi-axis signals into a single "picture quality" for a symbol
- Discounts redundant signals (e.g. golden cross + trend alignment both measure
  the same thing) so correlated evidence isn't double-counted
- Produces an ADR-0007-safe output: evidence descriptions ("strong bullish
  picture"), never buy/sell directives
- Supports a closed loop: slot fires are logged and later resolved to
  confirmed/false-alarm by measuring whether price followed through

## Decision

Build a **34-slot Confluence Signal Engine** as a first-class module
(`app/server/src/confluence/`).

### 1. Slot catalog (34 slots, 6 families)

Each slot is an independently-evaluable check whose firing state contributes
bullish or bearish evidence. Families: technical (15), institutional (5),
macro (5), seasonal (5), flows (3), sentiment (1). Slots carry an ADR-safe
`explain` note (evidence sentence, never a directive). `SlotBody`: bull / bear
/ exit (exit = bear evidence for an existing position).

### 2. Redundancy-aware rack evaluation

Slots are bucketed into redundancy groups (e.g. goldenCross +
trendAlignment + pullbackToEMA21). Within each group, evidence decays
geometrically (1 + 0.5 + 0.25 ...) so correlated signals count once, not
three times. The rack then labels the picture using evidence totals:

- `MIN_TOTAL_EVIDENCE = 1.0` (below → sparse)
- `DIRECTION_RATIO = 0.6` (bull/totals must reach this for bullish label)
- `STRONG_EVIDENCE = 4.0`, `MODERATE_EVIDENCE = 2.0` (magnitude thresholds)

Quality labels: strong/moderate/weak-bullish, mixed, weak/moderate/strong-
bearish, sparse.

### 3. CandleProvider seam (data abstraction)

Confluence evaluators resolve candles through a `CandleProvider` interface,
not `cache.get` inline. The cache-backed implementation folds in the freshest
live quote as a partial daily bar so mid-session evaluations see the current
price, not just the last EOD close. Weekly slots (50/200 cross, trend
alignment) use the weekly cache key. This seam is pluggable for future
replay/realtime sources.

### 4. Closed-loop signal history

Every slot fire is logged to `confluence_signal_history` with the as-of date,
rack, and picture quality at the time. A resolver later checks whether price
moved the expected direction over 4 weeks (bull slots → price up, bear/exit
slots → price down). A small dead-band (0.5%) treats flat outcomes as
unresolved rather than false alarms. Per-slot reliability weights (0.5–1.25)
allow the rack to self-tune over time.

### 5. Picture-change alert producer

The `confluence_change` alert type fires when the picture quality tier
changes (improved / deteriorated) or net evidence shifts beyond 0.35.
Throttled to 5 per hour per type. ADR-0007 framing: "the picture has
changed," never "act now."

### 6. COT data adapter

The CFTC Traders-in-Financial-Futures report (leveraged-funds long/short +
open interest) is fetched from CFTC's annual zip files and parsed. The
`cotPositioning` slot consumes this data. Registered under the `cftc`
vendor family with 1.5s min-interval pacing.

## Consequences

- The tRPC `confluence.*` router exposes slots, racks, evaluation, backtest,
  scorecard, and saveRack. The frontend `/confluence` page shows picture
  quality, per-slot evidence, and reliability scorecard.
- Three system rack presets are seeded on startup: Full Confluence (all 34),
  Technical Momentum (15), Macro + Flows + Sentiment (14).
- The 15-symbol confluence universe (PLTR, NVDA, AMD, AAPL, MSFT, SMH, XOM,
  JPM, UNH, COST, AMZN, CAT, LMT, LIN, NEE) + SPY benchmark are pinned
  into the demand set on startup.
- No advisory output is produced. All picture-quality labels describe
  evidence; they never recommend action.
- The closed loop is not self-executing: the `confluence.eval` procedure
  must be triggered to produce evaluations. The backtest procedure is
  read-only and query-driven. Future work can schedule periodic evaluation.