NeuPortal AI volatility forecasts put audit trail ahead of price direction

NeuPortal AI volatility forecasts should be judged by how well they estimate the size of market moves, rather than whether they predict direction, according to Andrii Bilous, co-founder of the Berlin-based firm. In a Finextra community post, Bilous argued that volatility estimates are more useful for risk teams and easier to audit than price targets in liquid markets.

Bilous said directional calls are difficult to score fairly because widely traded markets tend to absorb available information quickly. Volatility, by contrast, tends to cluster: quiet periods often follow quiet periods, while a shock can increase the probability of larger moves soon after. That persistence gives machine-learning systems a pattern to estimate, according to his post.

The argument has direct relevance for trading firms, exchanges, lenders and supervisors that rely on risk limits rather than market calls. A forward volatility range can feed into position sizing, margin settings, liquidation thresholds, value-at-risk calculations, conditional tail estimates and stress testing, Bilous wrote.

How does NeuPortal verify AI volatility forecasts?

Bilous said NeuPortal records each forecast before publication, hashes it with SHA-256 and anchors it to the Bitcoin blockchain through OpenTimestamps. SHA-256 creates a fixed digital fingerprint of a file or forecast, while OpenTimestamps can prove that the fingerprint existed before a later market outcome.

The method does not demonstrate that a forecast is accurate. It establishes timing, which is central to model-risk review because it separates a live prediction from a result selected after the event. Bilous said NeuPortal then scores outcomes publicly, including missed forecasts, against a stated scoring rule.

The firm uses Binance spot as the reference market for the experiment, according to Bilous. He said the public record is available at neuportal.ai/experiment.

Why does the method differ from a standard volatility band?

Bilous criticised the common practice of scaling a single volatility estimate by the square root of time, often described as sigma times root-t. That shortcut is widely used to translate short-horizon volatility into longer bands, but he said it can misstate the distribution for crypto and other assets with fat tails.

According to Bilous, NeuPortal compared that approach with the full Binance history for Bitcoin. He said the ratio of an empirically measured 80% band to a sigma-root-t band was about 0.80 at one day and rose to roughly 1.00 by 30 days. In his interpretation, the shortcut was too wide at short horizons and approximately right at longer horizons.

He also cited daily kurtosis near 16 for Bitcoin, placing much of the fat-tail effect in the extreme tails rather than the middle of the distribution. NeuPortal therefore builds bands from empirical quantiles of historical moves over the same horizon, with only a momentum-based adjustment to the median when a trend condition is met, Bilous wrote.

Bilous framed the work as a risk and governance exercise, not a claim to beat the market. He said better volatility estimates can reduce the cost of being wrong by improving sizing and loss bounds, while offering no reliable indication of where a liquid market will trade next.

This story draws on original reporting from Finextra Research.