Track 4

ML for Trading

Karpathy-inspired: micrograd → MLP → attention — applied to price data

Length

2.5h · 6 quests

Progress

0/6

Inspired by karpathy/nn-zero-to-hero, adapted to crypto. Build a one-neuron classifier from scratch, then an MLP for oversold-bounce prediction, then a tiny attention module. Ends with the hard lessons: distribution shift, leakage, and why your backtest Sharpe of 3.0 turns into 0.4 live.

Prerequisites:Algorithmic Trading

01What ML can — and cannot — do for tradingpython

Before you wire a neural net to your PnL, know the failure modes. This lesson saves you months.

10min

02micrograd: one neuron, from scratchpython

Karpathy's micrograd is 150 lines that teach you every line of PyTorch. We adapt it to predict next-bar direction.

18min

03A tiny MLP: oversold bounce classifierpython

Stack three neurons + a hidden layer. Train to classify 'this oversold bar will bounce >1%' vs 'it won't'.

20min

04Feature engineering: what to feed the modelpython

90% of ML-for-trading quality is features. Learn what works and what leaks the future.

15min

05A tiny transformer: attention over bar sequencespython

Karpathy's makemore-style transformer, adapted to 'what comes next after these 20 bars?'

25min

06Deploying ML to an agent — what goes wrongpython

You trained a model. Sharpe 2.5 on test set. Now what? This quest is the gap between lab and live.

15min