← Back to IRIS
IRIS Attribution

How your moves are explained

Every red and green number in your portfolio is decomposed into understandable factors and specific events. This page explains exactly how — every input to every score is here.

01How attribution works

Each holding's daily move runs through three tiers in sequence. Whatever a tier can't explain passes to the next.

Tier 1 — Factor decomposition. Your move is split into what the broad market did (your beta × the index), what your sector did apart from the market, and style factors like interest rates. What's left is the idiosyncratic (stock-specific) move — the part that isn't just the tide.

Tier 2 — Knowledge graph. The stock-specific move is matched against events traced through our supply-chain and competitive graph to your holding — including second-degree links, shown as context only.

What we have measured about competitors. Across 2019–2026, when a large S&P 500 stock made one of the day's biggest moves, the competitors in our graph made a 2%+ move in the same direction about 1.8 times as often as the average S&P stock (95% CI 1.7–1.9x), ranging from 1.4x to 2.3x by year. Same-day co-movement, not a forecast; stocks in the same industry co-move at a similar rate, and our competitor list is drawn from today's graph. After a big earnings reaction (5%+), a company's critical competitors made an opposite-direction 2%+ move only about one-third as often as a typical stock (0.32x, 95% CI 0.29–0.36x); rivals usually move with the news, not against it.

Tier 3 — Search. When the graph doesn't cover a significant stock-specific move, a bounded real-time news search looks for the catalyst — with the same evidence bar. When nothing clears it, we say so: an honest "no catalyst found" beats a plausible guess.

02Connection Strength (CS)

CS (1–10) answers: how confident are we that this causal link is real? It is additive: CS = relationship tier + revenue exposure + data-source quality + historical validation.

Relationship tier (the base): a direct, filing-confirmed dependency scores highest; a third-degree or thematic link scores lowest. Revenue exposure raises the score when we can quantify how much of your holding's revenue is at stake. Data-source quality rewards regulatory filings over news aggregators. Historical validation is coming soon (see §07).

In the product you'll usually see CS as words, not numbers — documented link (8+), strong link, plausible link, and possible link — not confirmed (below 4.5). That last band is exactly the confidence level at which every claim is hedged in prose, by contract. The raw score and its full breakdown are always one tap away.

03Explanatory Share (ES)

ES answers a different question: how much of the move does this event account for? We allocate the stock-specific move across competing events by three rules:

Temporal alignment — an event that happened after the price moved cannot be its cause, so it gets zero. Magnitude plausibility — an event is capped at a multiple of its fundamental (revenue) exposure, so a narratively-compelling but financially-minor link can't dominate. Relative confidence — higher-CS events get a larger share.

Market, sector, factors, events, and residual always sum to 100% of the move.

04The residual

The residual is whatever no factor or event explains. It is never dressed up: it renders as a labeled share of the move — thin trading where that's measurable, otherwise an honest "no specific catalyst identified." We don't publish a residual target; we publish the residual.

05Data sources

The knowledge graph is built from SEC filings, earnings calls, analyst and industry research, and institutional fund flows. Sources are tiered: primary regulatory (SEC, exchange) rank highest; major financial press next; general aggregators lowest. Prices are end-of-day, corporate-action-adjusted (splits and dividends are baked in, so they never masquerade as a move).

06Search attribution

When the graph doesn't cover a move, we search the live web and news. Search-sourced attributions are always tagged as such, must actually be about the company (subject-verified against the article, not just keyword-matched), and are scored more conservatively than graph links with capped share of the explanation — a news article is qualitatively different evidence from an SEC filing. Every gap search fills is also flagged to improve the graph over time.

07What we refuse to do

The scoring is only half the system; the other half is discipline enforced in code, not editorial policy:

Every causal sentence must cite its evidence. Generated prose that references a cause outside the scored evidence, fabricates a figure, or dodges a required hedge is discarded before you see it, in favor of a plain deterministic rendering.

Peers arbitrate cascade stories. Before a cross-company story stands, the implied peers' actual sector-adjusted moves are checked. A story the peers refute is confidence-capped and presented hedged — "the tempting story did not survive the peer check."

"Just the market" is a first-class answer. When a move tracks the market and sector, we say exactly that; when a move is stock-specific and nothing clears the evidence bar, we say that too. We don't invent stories to fill the gap.

Never predictive, never advice. Every sentence is past-tense and descriptive; forward-looking and recommendation language is rejected mechanically.

08Limitations

Graph coverage is deepest for large US companies; thinner for micro-caps, bonds, and foreign listings, where more of the move falls to search or residual. We hold no model of market expectations (consensus estimates, implied moves), so "no visible catalyst" means none in our data — not that the market had no reason. Historical-validation calibration (weighting scores by how well similar past attributions predicted forward returns) is logged from day one but not yet activated. Beta for holdings with little price history is estimated from the sector average.

09Version history

iris-1.1.0 — orthogonalized factor decomposition; direct (own-company) events; abnormality trigger raised to 1.5σ; cascade-relevance caps; direction conflicts kept-but-hedged; peer confirmation; hardened search attribution.

iris-1.0.1 — factor decomposition + knowledge-graph CS/ES + search fallback. Bounded ES, economic trigger floor, holding-direction revenue exposure.

iris-1.0.0 — initial three-tier engine.