Common Risk Factors in Decentralized AI Subnets

Abstract

I derive a size premium from the constant-product automated market maker used to price Bittensor subnet tokens and test the prediction using daily data on 128 subnets. A small-minus-big factor earns 1.01% daily (Newey-West t = 3.28). The December 2025 halving of token emissions, which the theory predicts should halve the premium, reduces it from 1.17% to 0.51% (p = 0.044). Exact slippage calculations show the premium is implementable only below \$10K in assets under management; at \$100K, transaction costs exceed gross returns.

Figures (11)

• Figure 1: Cumulative Returns of Factor Portfolios
• Figure 2: Cumulative Returns of Size-Sorted Tercile Portfolios
• Figure 3: Cumulative Returns of Momentum-Sorted Tercile Portfolios
• Figure 4: Factor Correlation Matrix
• Figure 5: Factor Sharpe Ratios and $t$-Statistics