This is the complete written course behind the Kalshi trading bot I run on real money — published post by post on Rebel Edge, now collected in one place.

Five posts. Every layer of the system. Code included.

What the course covers

Post 1 — Architecture
Why one model isn't enough. The dual-confirmation filter that rejects 91.9% of signals. How Chronos-T5-Small and LightGBM work independently and why they have to agree before a trade fires. Kelly Criterion position sizing. The full trading loop in 25 lines of Python.

Post 2 — Data Pipeline
Pulling live Binance OHLCV data with ccxt (no API key needed). Engineering 26 features across 8 categories with pandas-ta. Sine/cosine time encoding. The 3 live-injected Kalshi features that link the Binance chart to the prediction market. SQLite logging schema with 14 columns.

Post 3 — Training the Models
Setting up LightGBM binary classification. Why closed trade outcomes get 3× sample weight in training. How Chronos-T5-Small generates 100 probabilistic forecasts on CPU — no GPU needed. The dual-trigger retraining loop: every 6 hours on the clock and every 20 closed trades. Joblib persistence so restarts don't lose the model. Brier score and log loss validation before the new model goes live.

Post 4 — Paper Trading Validation
The one-flag architecture that makes paper mode and live mode identical. Five failure categories that paper trading catches and backtesting never will. The five SQLite tables that make go/no-go analysis possible — including the skipped_signals table most builders skip. Confidence interval math on 241 paper trades: what 51.5% actually means statistically. The three conditions the bot had to meet before going live.

Post 5 — Deploy, Monitor, Discipline
Setting up a Hetzner CX22 VPS for €3.79/month. The cron PATH gotcha that kills most first deployments. Systemd as crash recovery layer (not the scheduler). Rotating logs capped at 50MB. Telegram trade alerts. Retry logic for idempotent vs. non-idempotent API calls. Pre-written trading rules committed before the first live trade — because the version of you who wrote the rules before any losses is the only version worth listening to.

Who this is for

Python developers who want to build a real algorithmic trading system on a CFTC-regulated exchange — not a backtest, not a simulation, but a bot running on live capital with real market data, real API calls, and real settlement. You should be comfortable with Python. Familiarity with pandas and basic ML concepts helps but is not required — every concept is explained from first principles.

Live results from the bot this course builds

$428 portfolio · +18.95% from $360 starting balance · 132 live trades · 55.3% win rate · XRP contracts at 79%. Updated live on the Rebel Edge dashboard on Substack.

Not financial advice. Past performance does not guarantee future results.