MANTIS
SN123Rewards AI miners for embeddings that improve financial forecasting accuracy
Miners don't predict prices. They submit embeddings, learned numerical representations of market data, and validators measure whether those embeddings actually have predictive power for future price movements. It's a competition to produce the most informative market representations, not the most accurate forecasts.
// Embeddings that predict price.
MANTIS is a subnet where miners compete to create the best mathematical representations (embeddings) of financial data. These embeddings encode patterns in market data into dense numerical vectors. Validators then measure whether these embeddings contain genuine predictive signal for future price movements by computing "salience," how much each miner's embeddings improve forecasting accuracy.
The simple version: Imagine asking a thousand analysts to each summarize what they see in market data as a list of numbers. You don't ask them to predict anything. Instead, you test whose summary is most useful for making predictions. The analyst whose summaries consistently help the most wins. MANTIS does this with AI-generated embeddings.
Centralized equivalent: Think of quantitative trading firms' "alpha signals," the features they extract from market data to feed their trading models. MANTIS crowd-sources feature discovery through competition.
How it works:
- Miners submit encrypted embedding payloads for specified tickers at each sampled block. Embeddings have a defined dimension and are aligned with price data for evaluation.
- Validators collect encrypted payloads, use Drand beacon randomness for fair sampling, align embeddings with historical prices, compute per-hotkey salience using multi-salience models, and set on-chain weights based on predictive contribution.
- The problem it solves: Quantitative trading firms spend millions developing proprietary features (signals) from market data. The search for alpha is expensive, secretive, and concentrated among a few elite firms.
- The opportunity: If decentralized feature discovery can produce embeddings with genuine predictive power, it democratizes a process currently monopolized by Renaissance Technologies, Citadel, and similar firms.
- The Bittensor advantage: Encrypted submissions prevent information leakage. Drand beacon randomness ensures fair evaluation. The competition rewards information content, not model complexity, meaning a simple but insightful embedding beats a complex but noisy one.
- Traction signals: 124 commits across 5 contributors with 4-12 commits per week. 244 active miners. Encryption and Drand-based fairness. 1,451 holders.
Category: Financial Forecasting and Trading Signals | Centralized Competitor: Numerai, WorldQuant BRAIN, Two Sigma
MANTIS takes a subtly different approach from most financial subnets. Rather than asking miners to predict prices (which is what most do), it asks them to produce representations that contain predictive information. This distinction matters: embeddings are more flexible than forecasts. They can be combined, reweighted, and fed into different downstream models.
Mechanism:
The evaluation pipeline is sophisticated. Validators use Drand beacon randomness to sample blocks fairly (preventing miners from gaming the evaluation timing). Encrypted payloads prevent front-running and information theft. Once collected, validators align embeddings with actual price movements and compute multi-salience scores, measuring how much each miner's embeddings reduce prediction error.
The salience scoring is the key innovation. Rather than rewarding the single best predictor, MANTIS rewards miners whose embeddings add the most information when combined with others. This encourages diverse approaches: a miner capturing momentum signals and a miner capturing sentiment signals can both score well because their embeddings are complementary.
The codebase has 124 commits across 5 contributors, with recent work on trading strategy implementations (hit-first, range-breakout). The primary contributor is Barbariandev. 244 active miners is one of the highest counts in our coverage.
Market metrics are early-stage. At 15,386 TAO market cap with 1,451 holders, MANTIS is small. Root proportion of 0.234 is one of the highest we've seen, indicating the subnet is still establishing organic demand. Gini of 0.701 and HHI of 0.111 show concentrated ownership.
Net 7-day flow is essentially flat at 3 TAO. The 30-day return of 9.4% is modest but positive.
- High root proportion: 0.234 is one of the highest in our coverage, meaning the subnet relies more on protocol subsidy than organic demand.
- Concentrated holdings: Gini of 0.701 and HHI of 0.111 indicate significant concentration.
- Unproven alpha: Whether the embeddings produced by MANTIS miners actually contain trading alpha remains to be demonstrated with live results.
- Financial regulatory risk: Producing financial signals through a decentralized network may attract regulatory scrutiny depending on jurisdiction.