Probabilistic forecasting returns a distribution of possible outcomes with calibrated probabilities. For markets, the output is thousands of plausible price paths over a horizon.
Why it matters for markets
Markets are non-stationary. The same input produces very different outcomes depending on regime, liquidity, and tail conditions. A point forecast hides this. A distribution makes it the message, and the substrate for honest risk work:
- Probability of a 5% drawdown over the next 4 hours.
- Portfolio loss under correlated stress.
- Agent decisions weighted by counterfactual scenarios.
- Calibrated tail risk that does not collapse to the median in regime transitions.
What makes a forecast calibrated
A calibrated model is one where stated probabilities match observed frequencies. If the model says "30% chance of a -2% close in 4h", that event should happen ~30% of the time across many forecasts. Calibration is testable and scoreable.
- Cross-entropy is the proper scoring rule. Lower is better.
- Extropy is the normalized inverse of contextualized cross-entropy. Higher is better. Used as the primary head-to-head metric at duonlabs.com/arena.
How Duon Labs operationalizes it
Voyons takes raw market data and emits thousands of calibrated scenario paths per request. The output is a probability surface: regime dynamics, tail risks, directional confidence. REST, Python SDK, and pay-per-request via x402 for autonomous agents. Every release competes head-to-head in the public arena. Full loop at live.duonlabs.com.
Questions
+ How is probabilistic forecasting different from a confidence interval?
A confidence interval is a summary of a distribution. A probabilistic forecast keeps the full shape, including tails. You can always reduce a distribution to an interval. You cannot reconstruct one from an interval.
+ How is probabilistic forecasting different from an LLM applied to markets?
LLMs are optimized for text. Markets are numerical, temporal, non-stationary. Voyons learns directly from raw market data and outputs probability distributions. It never hallucinates numbers because numbers are its native output type.
+ Can probabilistic forecasts be wrong?
Individual events can land in low-probability regions. That is expected. The right measure of quality is calibration: do stated 30% events actually happen ~30% of the time? Evaluated continuously via extropy in the public arena.