Epsilon: scoring idiosyncratic risk for IC review

Epsilon is the engine in our portfolio that scores idiosyncratic risk — the issuer-specific risk that systematic factor models miss. It is also the product we built first, and the one we offer in a free tier, because we believe the idiosyncratic lens is the single most useful addition you can make to a factor-model-based risk view.

This post is a product spotlight: what Epsilon does, how it scores, what the evidence chain looks like, and how to put it to work inside an IC workflow.

8-minute read · Updated 16 May 2026

Key takeaways

• Epsilon scores idiosyncratic risk on a 0–1 scale using three independent sub-scores: time-series anomaly, narrative coherence, and microstructure signal.
• The two-of-three coherence rule suppresses single-signal noise — a flag requires at least two sub-scores elevated plus a top-decile composite.
• Three production use cases: pre-IC name screening, hedging trigger discipline, and diligence on prospective additions.
• The free tier covers the full investable universe with anonymised flags and basic evidence chain. Paid tiers add named issuers, custom universe, and PortIQ integration.

The problem Epsilon was built to solve

Factor models have served the industry well for forty years. They explain a large share of cross-sectional return variation, and they let a portfolio manager think about risk in compact, intelligible buckets — market, size, value, momentum, quality, low-volatility, and so on.

What factor models do not explain well is the residual: the part of an issuer's risk that is specific to that issuer. The dot-com infrastructure unwind, the 2008 mono-line shock, the regional bank failures of 2023, and the more recent AI-capex re-rating events are all primarily idiosyncratic events. Looking at them through a pure factor lens, the warning signs are diffuse and easy to miss.

Epsilon is the lens for that residual. It scores every issuer in the investable universe on a 0–1 scale of idiosyncratic risk, daily, with an evidence chain attached.

How the score is constructed

Epsilon's composite score is built from three independent sub-scores. The architecture is deliberately conservative: we suppress single-signal noise by requiring coherence across at least two of the three sub-scores before a flag is surfaced.

Sub-score 1 — Time-series anomaly

Drawn from our Time-Series Forecasting engine. Captures unusual price behaviour relative to the issuer's own history and relative to peers, controlling for systematic factor moves. Flags include unusual implied volatility relative to realised, unusual moves in idiosyncratic residuals, and shifts in autocorrelation structure.

Sub-score 2 — Narrative coherence

Drawn from the Risk Brain (our LLM-based agent layer working on filings, transcripts, and news). Captures shifts in management language, sentiment compression, the introduction or removal of specific risk disclosures, and the alignment (or misalignment) of management language with sell-side and rating-agency commentary.

Sub-score 3 — Microstructure signal

Drawn from Microstructure Watcher. Captures bid-ask widening, order book depth changes, trade frequency anomalies, and unusual cross-market correlations.

The composite score is a coherence-weighted combination. An issuer scoring high on only one sub-score does not generate a flag. Two-of-three with a top-decile composite does.

What the output looks like

A typical Epsilon view for a single issuer shows:

• The composite score (0–1) and its trajectory over the last 90 days.
• Each sub-score, colour-coded, with the specific underlying signals that drove it.
• The evidence chain — links to the specific filings, news items, and market microstructure observations that contributed.
• Peer comparison — where this issuer sits relative to its sector and size cohort.
• Suggested actions — calibrated to the magnitude and persistence of the score.

For a portfolio view, Epsilon aggregates: which positions sit in the top-decile of idiosyncratic risk today, which clusters of names are flagged together, and which sector / region cuts have the highest aggregate Epsilon exposure.

How to use Epsilon inside an IC workflow

We have seen Epsilon used well in three distinct ways:

Use 1 — Pre-IC name screening

The portfolio team runs Epsilon as an overlay on the current book the day before an IC. Names with newly elevated scores get a short Risk Brain summary attached. The IC starts with a coherent, ranked list of names that warrant discussion rather than the usual top-of-mind list.

This is the use case Epsilon was built for. It compresses what used to be a manual sweep — and which often missed names that had not surfaced through conventional channels — into a five-minute review.

Use 2 — Hedging trigger discipline

Some clients pre-define hedging actions that activate when an Epsilon composite crosses a threshold. The rule is straightforward — "if Epsilon composite > 0.75 for ten consecutive sessions on a name representing > 2% of book, the pre-staged hedge activates" — and the discipline removes the temptation to second-guess in real time.

The pre-staging is what matters; the hedge is already designed, costed, and approved before the trigger fires.

Use 3 — Diligence on prospective additions

Before a new issuer is added to the book, Epsilon's full history on that issuer goes into the diligence file. A name with elevated idiosyncratic scores in the trailing six months does not necessarily get rejected — but it does get added with a specific risk plan rather than a generic one.

The five questions to ask in your Epsilon trial

When you start a free-tier evaluation, put the engine through these tests:

1. Backtest against your blow-ups. Run Epsilon's history against the three or four single-name events that hurt you most over the last five years. We will tell you what we would have flagged, and when.
2. Force a false positive. Pick a name you are convinced is fine and see how Epsilon treats it. The defence against false positives is the two-of-three rule; the defence against false negatives is the sub-score visibility. Both should be inspectable.
3. Test the evidence chain. Click through from any flagged name back to the underlying filings, news, and microstructure observations. We promise lineage in one click.
4. Look for coherence with your own analyst views. The Epsilon score should sometimes confirm and sometimes contradict your analysts. When it contradicts, the reason it contradicts is more useful than the score itself.
5. Push the universe. If a name in your investable universe is not covered in the free tier, ask. Many of our paid-tier expansions started as a trial-user request.

Where Epsilon fits in the broader portfolio

Epsilon is one of the eight Quantitative Engines in our stack — alongside the Factor Model, VaR/CVaR, Portfolio Optimisation, Time-Series Forecasting, Apex Intelligence, Signal Generation, Commodity Intelligence, and Vantage Valuation. All eight are available as REST APIs (OEM/white-label) for clients who want to embed them. PortIQ and CyronOS bring them together with the agent layer and a full IC workflow.

But you do not need any of that to start. Activate Epsilon in the free tier, run it against the book you have now, and see what surfaces.

Frequently asked questions

How is idiosyncratic risk different from systematic risk?

Systematic risk is the portion of an issuer's risk explained by common factors (market, size, value, momentum, etc.). Idiosyncratic risk is the residual — issuer-specific risk that factor models treat as noise. Idiosyncratic risk drives the most consequential single-name events.

What is the two-of-three coherence rule in Epsilon?

Epsilon scores three independent sub-scores per issuer. A flag is only surfaced when at least two of the three sub-scores are elevated and the composite ranks in the top decile of the universe that week. This filters single-signal noise.

Is Epsilon really free?

Yes. The free tier covers the standard investable universe, daily refresh, the composite and sub-score breakdown, basic evidence chain, and 90 days of history. Paid tiers extend depth, universe, and platform integration.

Can I integrate Epsilon with my existing risk system?

Yes. Epsilon is available as a REST API for OEM and white-label deployment, and it integrates natively with PortIQ, CyronOS, and the Risk Signal Engine BYOD.

How does Epsilon avoid false positives?

The two-of-three coherence rule is the primary defence. The sub-scores themselves are calibrated weekly against a labelled validation set. Persistent false-positive patterns drive sub-score retraining.