AlphaTerm

A one-person AI trading desk: talk to the AI to check markets, write strategies, run backtests, and place orders—three permission tiers determine how much control you hand over.

What this is

A Bloomberg-terminal-style, single-user trading interface. The engine is the mature open-source quant framework Freqtrade, and the AI sits alongside as a researcher and trading assistant: you talk; it pulls data, writes strategies, backtests, and places orders within the scope you authorize. No multi-user, no SaaS.

What problem it solves

Quant frameworks can execute strategies automatically, but the layer of "researching markets, tuning strategies, and changing configuration" still depends entirely on humans writing code. Empirical evidence also shows that LLMs lose money when making per-trade buy/sell decisions (in Alpha Arena's live trading with six large models, four lost money); what they are good at is exactly the research and configuration layer. So the division of labor: AI handles research and strategy, the engine handles execution, and the human decides how much control to hand over.

Use cases

You type "Write an EMA crossover strategy and backtest it" in the chat box. The AI writes the file, passes a safety check, runs the backtest in an isolated container, and reports 664 trades, a 22.4% win rate, and a −0.43% return. It fails, so the AI iterates three more versions on its own. Another strategy that clears the threshold (drawdown ≤20%, at least 10 trades, positive net return) goes live automatically.

Highlights

• Three permission tiers: read-only can only view; confirm turns every action into a card and waits for your approval (default); bypass is fully automatic. The tier restricts the AI, not the human's decision-making power.
• Every action goes into the audit trail: in bypass mode the AI can change risk-control parameters, but every modification is written to the audit log and highlighted in the dashboard; this chain cannot be turned off.
• What got exposed in real-world testing has been fixed: when the AI said "add ETH" it silently removed BTC, so the whitelist tool was changed to a strict choice among add/remove/replace.

Where it stands now

71 commits in three days, and all three phases were verified end-to-end with real API keys. Right now it is running a real strategy locally in Docker with real market data and fake money; out of 38,000 lines of code, half are tests. Live trading belongs to Phase 4, with a review checklist before touching real money. For personal use; a demo site is available, with simulated data.