Jev on Base

We gave an AI trader 40,000 calls on Base. It did worse than a coin flip.

Viral posts show Jev, a fast decision model, trading crypto every block. None of them checked whether its calls are right. We did: pre-registered, 90 days, four markets on Base, every decision scored against what the price did next.

How well Jev's probability that the price rises in 15 minutes separates rises from falls (AUC). 0.5 is a coin flip. Bars are 95% intervals.
Data table

What the viral posts actually claim

The posts sell speed and cost. We went looking for the profit.

SourceWhat it showsProfit shown

Does Jev know where the price is going?

No. In every market and horizon its probabilities rank outcomes about as well as chance, and usually a little worse. Free rules do as well or better.

Market
Skill by horizon
Jev against two free rules: momentum bets the last move continues, mean reversion bets it reverses.
Data table
Calibration at 15 minutes
Data table

One decision, as Jev saw it

State sent to Jev


          

          

            
Jev's answer: probability the price is higher after

            

            

            
The 15-minute question

            

          

        

      


      

        How we measured it
        
    
      
    
    
    

    
    
      
    Then what is it doing?
    
      
    Following the trend. Its probability of a rise tracks the last 15 to 60 minutes of price and order flow. On Base, pool prices tend to reverse over that horizon, so following the trend loses.
    
      
    
        
    
          
    
          
          
    Rank correlation between Jev's probability and each input. Near 1: Jev moves with that input.
    
          
    Data table
    
        
    
        
    
          
    Jev tracks momentum, month after month
    
          
          
    
          
    Data table
    
        
    
      
    
      
    
    
    

    
    
      
    Did we get something wrong?
    
      
    We tried to. The pipeline finds skill when it exists, the result survives a change of price source, and two independent projects found the same pattern.
    
      
    
    

    
    
      
    What if you trade it anyway, or do the opposite?
    
      
    Trading it loses to holding. Flipping it gives a weaker copy of a free linear model, still far from paying for its trades.
    
      
    
        
    
          
    
          
          
    
        
    
        
    
          
    Out of sample: the last 45 days
    
          
          
    Every model was fit or tuned on the first 45 days and scored on the last 45.
    
          
    Data table
    
        
    
      
    
      
    
        
    The gap no prompt can close
    
        
        
    Each trade on a Base pool pays the pool fee, price impact and gas both ways. Given how far prices move in 15 minutes, breaking even takes the hit rate on the right. The best signal we found reaches the one on the left.
    
      
    
    
    

    
    
      
    Is there money around it without predicting anything?
    
      
    Yes, but it belongs to specialists who were already there. We checked three places and Jev adds nothing to any of them.
    
      
    
    
    

    
    
      
    Where Jev might still help
    
      
    Jev is cheap and quick at judging text. That turns into an edge only where text is the bottleneck, nobody else is reading it fast, and the price is slow to react. In liquid crypto those three rarely meet. The candidates left are narrow: long-tail tokens reacting to unstructured posts, filtering rugs at launch, and reading prediction-market rules. None is tested here.
    
      
    Test the window before the model. For example, first measure how long a pool takes to reprice after a listing announcement. If it is one block, bots already own it and no classifier will help.
    
    
    

    
    
      
    Run it yourself
    
      
    Everything is in the repository: the pre-registered protocol, the code, and the reports. You need Node 23.6 or later and a Vercel AI Gateway key. The public Base RPC works, just slowly.
    
      
    git clone https://github.com/r4topunk/jev-trading-study
cd jev-trading-study && pnpm install
node scripts/backfill.ts --days 90
node scripts/backtest.ts --market eth --days 90 --every 15 --max-usd 0.8
node scripts/compare.ts
node scripts/audit/positive.ts