Allora Network’s Forge Expands with Volatility Topics, Enabling a New Class of Prediction Models

Four assets, one horizon, and Allora Network's first prediction family beyond price.

Introducing Volatility Topics

Four realized volatility prediction topics are now live on Allora Forge. Topics 20 through 23 forecast 15-minute realized volatility for BTC/USD, ETH/USD, XRP/USD, and SOL/USD.

These are the network's first mainnet volatility topics and its first expansion beyond price and log-return prediction into a second prediction family. Builders can now develop and evaluate models against a target that measures the magnitude of market movement rather than its direction.

What launched

Topics - 20 (BTC), 21 (ETH), 22 (XRP), 23 (SOL), all versus USD

Prediction target - Realized volatility over the next 15 minutes

Cadence - New predictions collected every epoch, roughly every 5 minutes

Scoring - Predictions scored against realized volatility computed from market data after the 15-minute window closes. Rewards follow regret-based scoring

Network - Allora mainnet

Consumption - Topic inferences served through the Allora API, with per-topic streams available through the API gateway on a pay-for-inference basis

The four topics have shown consistent worker performance on testnet.

The prediction target

Realized volatility measures the magnitude of price movement over a defined window using observed market data. Unlike implied volatility, which is derived from option prices, realized volatility is computed from the prices that traded. These topics forecast the realized value for the next 15 minutes, producing an objective target once the window closes.

That target gives models a defined role in systems that condition on expected market activity. A volatility inference can inform position sizing, option pricing, market-making spreads, execution logic, hedging, and controls that behave differently across calm and turbulent regimes.

The forecast is not directional. It supplies context for deciding how much exposure a system should take, how wide its thresholds should be, or whether current conditions match the regime in which another model performs well.

How each topic runs on Forge

Each topic defines one target, one horizon, and one scoring loop. Participants contribute through distinct roles:

• Workers, also called inferers, operate models and submit predictions.
• Forecasters estimate which workers will be accurate under current conditions, sharpening the ensemble.
• Reputers bring off-chain ground truth and score submissions.
• The network inference is the combined, accuracy-weighted prediction consumed by an application.

The topics run continuously:

1. Every epoch, roughly every five minutes, registered workers submit forecasts for realized volatility over the next 15 minutes.
2. The network combines worker predictions and forecaster inputs into one network inference.
3. After the forecast window closes, realized volatility is computed from market data.
4. Submissions are evaluated through regret-based scoring, updating participant performance and reward allocation.

Because the prediction horizon is 15 minutes and the submission cadence is roughly five minutes, forecast windows overlap. Multiple forecasts can be at different stages of maturity at the same time, while the latest completed round supplies a resolved score.

The process is automatic. Performance is determined from predictions and measured ground truth rather than discretionary review.

Why the ensemble matters for volatility

Volatility regimes shift. A model calibrated to calm markets can degrade when market structure changes, while a model that performs during turbulent periods may contribute less during stable conditions.

Forge addresses that variability by combining independent worker predictions and using forecasters to estimate which workers are likely to be accurate under current conditions. Regret-based scoring evaluates how participant inputs affect the quality of the network inference. The output is one continuously evaluated inference rather than a collection of raw model predictions that every consumer must reconcile independently.

Running the same prediction family across BTC, ETH, XRP, and SOL also creates a comparative view at a common horizon. A multi-asset system can consume the four inferences together to observe which assets are entering more turbulent regimes and which remain relatively stable. This cross-section can inform allocation and hedging across a portfolio, not only decisions within one pair.

As additional assets and horizons come online, that cross-section can develop into a broader volatility surface for applications to consume.

Consuming the network inference

Each topic inference is served through the Allora API. Per-topic streams are available through the API gateway on a pay-for-inference basis, and each of the four topics can be queried independently.

A consuming system reads one accuracy-weighted network inference per topic per round. It can use one asset or combine all four, then pass the values into existing sizing, pricing, risk, or execution logic. Consumers do not need to aggregate the underlying worker outputs themselves.

Reading the topics as a group adds a relative signal. At a shared horizon, the four inferences show how expected movement differs across major assets. For multi-asset strategies, those differences can provide more useful regime context than a single-asset estimate in isolation.

Building for volatility topics

Realized volatility is a well-defined modeling target with established statistical and machine learning approaches. Builders working with GARCH-family models, realized measures derived from high-frequency data, or ML systems trained on market microstructure already have relevant foundations for experimentation.

The path to these topics runs through Forge. Builders develop a model and establish its performance on live testnet data. Models that prove themselves can then be promoted to mainnet topics. The workers now contributing to Topics 20 through 23 followed that progression before the topics launched on mainnet.

This creates a live evaluation environment for comparing modeling choices across repeated epochs and changing regimes. Builders can test feature sets, training windows, model families, and update strategies against the same target and horizon while building an observable performance record.

Extending the network beyond price

Volatility topics expand the range of prediction targets supported by Allora Network. Price and volatility serve different system requirements: one estimates direction or return, while the other estimates the magnitude of movement. Supporting both gives builders another class of models to develop and gives consumers a risk signal that can sit alongside directional forecasts.

More assets and prediction families are planned. Further announcements will follow when they are live.

Network outputs are aggregated predictions, not financial advice. No guarantee of accuracy or fitness for any purpose. Consumers bear all trading risk.

‍