Calibrated Confidence

Every engine emits a raw conviction (0–100) — an LLM's uncalibrated confidence in a directional move. That number isn't a probability. The calibration layer maps it to a calibrated confidence: the empirical chance the prediction comes true, learned from resolved outcomes.

On a token report or radar card, the calibrated value is what's published as the headline probability (with a confidence interval), while the raw conviction remains visible for transparency.

For each engine, a logistic regression is fit over its resolved predictions: P(hit) = sigmoid(intercept + slope × conviction/100). The fit runs on two horizons — a 7-day canonical headline model and a 2-day fast-feedback model that activates sooner.

Each model stores its coefficients, an empirical hit rate per conviction bucket (0–40, 40–60, 60–75, 75–90, 90+), an overall hit rate, and a Brier score (mean squared error of the probabilistic forecast — lower is better). A Brier near 0.25 is no better than a coin flip; closer to 0 is sharp. Fits are recency-weighted with a 60-day half-life, so probabilities track the current market regime rather than an all-time average — while the all-time hit rate is still reported honestly alongside.

A model fit in a Greed market and a model fit in a Fear market behave differently. Arvio fits regime-conditional models — separate fits for Fear, Neutral, Greed, and Extreme — so the regimes don't blur each other.

When a specific regime doesn't yet have enough resolved samples, the platform falls back to the regime-agnostic overall model. The token-analysis calibrator picks the regime-matching model first, then the overall, so a report's probability always reflects current conditions when the data allows.

Beyond the univariate map, a multivariate logistic regression conditions on discrete evidence features — risk tier, chain, market regime, signal type, and prediction direction. This learns which evidence is predictive and which is noise.

The output is a per-feature breakdown: the empirical hit rate and a marginal percentage-point contribution for each feature value. Those measured weights are fed back into the curation prompts — so the LLM credits reliable evidence and discounts evidence that historically underperforms.

Not every engine is best at the same timeframe. Arvio evaluates hit rate at 2, 7, 14, and 30 days, and picks the optimal horizon as the one with the greatest separation from a coin flip (|hit rate − 0.5|), with a minimum sample bar.

The 2-day model gives faster feedback while the 7-day base accumulates; the 7-day model is the canonical headline you see across the platform.

Markets change, and a model can go stale. A drift monitor compares each engine's recent realized hit rate against its fitted model's expected rate. When the gap crosses a threshold, a drift alert fires and the model is re-fit.

Alerts auto-resolve once performance realigns with the model. Active drift alerts surface on the track record so you can see when an engine is in flux.

A new engine has no resolved outcomes yet. Rather than publish nothing, a cold engine inherits the warmest donor engine's fitted model as a prior — so the Token Analysis engine, for example, can publish a calibrated probability before it has its own track record.

Once the engine's own outcomes cross the sample threshold, the transferred prior is replaced by a model fit on its own data.

Arvio mines empirical regularities from its resolved outcomes and stores them as self-improving intelligence objects of type "pattern". These aren't hand-written rules — they're patterns the platform discovered that historically preceded winning calls.

Active, high-conviction patterns are injected into curation as priors: tokens whose evidence matches a learned pattern are favored; tokens matching a pattern's invalidation criteria are avoided. The loop compounds.