MANTIS

Financial institutions and trading firms face a persistent challenge: sourcing reliable price predictions across multiple asset classes and forecasting horizons without building proprietary infrastructure or relying on single-vendor models. Centralized prediction services introduce latency, vendor lock-in, and opacity in how predictions are scored and weighted. Teams cannot easily benchmark their models against peer approaches, nor can they contribute predictions without surrendering algorithmic control.

A decentralized network of independent miners can address this by submitting encrypted predictions to validators who independently score them using consistent, transparent methodologies. Encryption ensures miners cannot observe competing predictions before submission; timelock cryptography prevents validators from peeking at payloads before a predetermined maturation window. Validators use walk-forward machine learning to score each miner's predictions on calibration, feature quality, and predictive power, then aggregate scores across challenges and set on-chain rewards proportional to demonstrated performance.

MANTIS (Bittensor Subnet 123) implements this model with 12 parallel prediction challenges spanning binary direction (1-hour horizons on ETH, forex pairs, and precious metals), volatility regime classification (17-dimensional quantile predictions on ETH and BTC at 1-hour and 6-hour horizons), barrier-hit timing (barrier-crossing prediction on ETH), multi-asset range breakouts (33 crypto assets), cross-sectional return ranking (33 assets, 4-hour horizon), funding-rate ranking (20 assets, 8-hour horizon), and capital-at-risk bracket trades (BTC, variable 1–336 hour regimes with collateral locking and significance-gated payout). Validators score miners using L2 logistic regression with challenge-specific feature selection and importance weighting, apply sybil resistance via coefficient regularization, and set on-chain weights weekly after EMA smoothing to reduce variance.

MANTIS (Bittensor Subnet 123) is a machine learning prediction market that scores miners on their ability to forecast price movements across multiple asset classes and prediction horizons. Validators collect encrypted miner embeddings, decrypt them after a timelock maturation window, and use walk-forward L2 logistic regression with challenge-specific scoring paths (binary classification, volatility regimes, cross-sectional ranking, barrier hits, range breakouts, and capital-at-risk bracket trades) to compute per-miner salience scores, which are aggregated by challenge weight and smoothed via EMA before being set on-chain as validator weights.