TradeBot Companion

What it is

TradeBot Companion is a local trading cockpit I built to run two cooperating crypto strategies against a paper account and watch them in real time. The point wasn’t a single clever model. It was owning the whole loop: pulling market data, engineering features, training and serving an ML classifier, executing against risk/budget guards, and surfacing all of it in a dashboard I can actually reason about (and talk to).

How it works

• Data + features. Market candles (OHLCV) are stored locally in SQLite per symbol/timeframe and turned into 15 engineered features — moving averages and their crossover, volatility, ATR, MACD histogram, RSI, Bollinger width, multi-window returns, and even API-latency signals as a data-quality guard.
• ML trader. A RandomForestClassifier (currently 300 trees trained on 3,444 bars) predicts trade direction per candle; a long is only taken when the forest’s vote share clears a 60% confidence threshold. The dashboard tracks accuracy, precision/recall, and feature importance from the live pickle, with per-symbol realized PnL.
• Volatility bot. A rules strategy that opens positions with explicit TP/SL brackets, gated by a risk guard that enforces a daily budget and exposure caps.
• Execution. Orders/fills run through the Luno exchange API in a paper account that starts at R10,000; the cockpit tracks the account’s journey, spend, and realized/unrealized PnL.
• Companion. A conversational layer answers plain-English questions over the live data and produces a coach-style session summary.

Inside the model

The newest part of the cockpit is a dedicated ML Model tab that turns the classifier from a black box into something you can actually inspect:

• The decision pipeline, drawn. A schematic walks the exact route a signal takes: OHLCV candles → 15 features → 300 trees → majority vote → confidence threshold → go long / skip. Everything on the page is read live from the trained model file, not hard-coded.
• Live decision. The latest candle runs through the forest on demand: a confidence gauge against the 60% threshold, every tree’s vote rendered as a square (e.g. 14 long / 286 short → SKIP), matching what the trader logged for the same bar.
• Tree explorer. Pick any of the 300 trees, expand its depth, and the SVG highlights the path the current candle takes to a leaf — with each split (bb_width ≤ 0.0083 → LEFT, …) logged step by step underneath.
• Plain-English features. Each of the 15 features gets a one-line explanation and an importance bar from the live pickle, so atr_pct at 12.3% means something to a reader who has never opened the code.
• Honest evaluation. A confusion matrix and ROC curve (AUC 0.572) on the held-out test split, captioned for what it is: a modest edge, reported honestly on unseen data.
• One-click retrain. A single button chains dataset build → training → refresh, streaming script output into the live log console; forest size, depth, and test split are tunable from the operator-gated Control Room.

Why it matters

It’s a compact demonstration of the full data-to-decision pipeline I care about: applied ML on engineered time-series features, model interpretability that holds up to scrutiny (tree paths, confusion matrices, honest AUC — not just an accuracy number), real-time systems design, robust guardrails around automated money decisions, and an interface (visual and conversational) that makes the system legible.

Next: a self-contained replay demo of the cockpit (canned data, no live infra) so this case study becomes interactive right here on the site.