Hey all!
EigenPhi just shared a small MEV-labeled DeFi dataset (see the Datasets thread
). I want to kick off some experiments with simple, beginner-friendly models before diving into fancy stuff—and I’d love your take on which one to start with.
Below are three baseline candidates I’m weighing. If you’ve used any of these on tabular/transactional data, please drop your war stories, code snippets, or gotchas.
| # | Model idea | Why it’s beginner-friendly | What I’m unsure about |
|---|---|---|---|
| 1. CatBoost (or XGBoost) | Straightforward tabular classifier. Handles numeric columns (volumes, gas, P&L) and one-hot for small categorical fields. | Great docs & minimal preprocessing. Best tricks for the wallet address column | hash buckets? label-encode? |
| 2. FT-Transformer (grain-of-salt “tabular transformer”) | PyTorch-based library; drop in a CSV and let it embed each feature. | Lets me test attention without building a monster pipeline. | Does it overfit fast on a few-hundred-row demo? |
| 3. Logistic Regression / Random Forest baseline | Good sanity check: “Can a super-simple model separate Arbitrage vs. SandwichVictim?” |
Easy to run, easy to interpret. | Worth engineering ratios (profit/gas, token count) first—or just raw features? |
Questions for the crowd
- Which of these would you pick for a first pass, and why?
- Got a Colab / notebook template you like?
- Any feature-engineering hacks for blockchain addresses or token symbols that don’t need deep crypto knowledge?
Thanks in advance for sharing—let’s help more folks jump into DeFi behavior modeling without getting lost in the weeds!