What ML can — and cannot — do for trading

Most ML-for-trading that students build: 1. Overfits the training set brutally 2. Shows amazing backtest Sharpe 3. Loses money in live trading 4. Gets abandoned

Why? Because financial data is low signal-to-noise + non-stationary. The regime changes and your model breaks.

• Pattern classification — "is this bar an oversold bounce candidate?"
• Regime detection — "is BTC in chop or trend right now?"
• Feature engineering — learning which indicator combinations matter
• Sizing decisions — calibrating a continuous confidence score

• Price prediction (pretty much a non-starter for chaotic markets)
• Small-sample learning (you almost never have enough non-overlapping data)
• Avoiding overfit without heavy regularization + out-of-sample testing
• Reading the news (sentiment models lag the actual move)

We'll follow Karpathy's nn-zero-to-hero:

• Build micrograd first — one neuron from scratch, backprop by hand
• Then a tiny MLP — classify oversold/not-oversold bars
• Finally a transformer-lite — attention over bar sequences

Each quest has exercises you'll run in the Pyodide sandbox below.

Answer: "Which ONE of the four ML use-cases above is agent-1's current v2.6 strategy already doing manually?" (rephrase as a hint: the agent uses RSI thresholds to classify oversold... something like that.)