v0.1.0: our own stock discovery, an LLM trade desk, protected positions and profit-taking

The system now finds its own stocks, argues every trade with LLM agents before it opens, keeps every position protected at the broker, and takes profit at +1R. It trades a paper account; no edge has been demonstrated yet (see Status).

Added

• Stock discovery (api/scout/). An hourly ranking of the stocks worth watching today:
  • Pool: Alpaca's most-active stocks and top gainers, the last day of news, Reddit mentions (ApeWisdom), StockTwits trending, and what the world's exchanges are trading.
  • Scoring: past-window performance, today's move, home-exchange move, news and discussion.
  • Every ranking is logged; python -m scout rank|scorecard|backtest evaluates them.
• World exchanges. The most-traded stocks on London, Xetra, Paris, Amsterdam, Zurich, Hong Kong, NSE, Tokyo, Korea and Taiwan, mapped by company name to US lines Alpaca can trade:
  • An SEC foreign-issuer check keeps same-named US companies out.
  • One watch slot per exchange.
  • Stocks with no US line, or only an OTC one, are shown with the reason.
• Watcher agent. Follows each pick every 5 s. The scout strategy trades a pick once the watcher's confidence has held above the bar.
• Trade desk (api/desk/). No buy goes out without approval:
  • Observer watches the signal for 90 s.
  • Analyst: qwen3.5:9b with reasoning, entirely on the GPU.
  • Critic: qwen3:4b, split between GPU and RAM.
  • Decision: the mean probability must beat the trade's breakeven plus a margin set by the risk dial.
  • The case file carries the user's risk appetite. The reasoning streams live to the dashboard. Every case is logged with its trade's outcome.
• News ingestion. Every headline is logged with its named entities, event keywords and scores.
• Profit-taking (api/engine/profit_manager.py):
  • At +1R, half is sold and the stop moves to breakeven.
  • The stop then trails the high by 1R.
• Broker-side protection. Entries are real bracket orders, and a check every 60 s gives any unprotected position an OCO stop and target.
• Spread monitor (api/feeds/spreads.py). A median of the consolidated (SIP) quotes from 16 minutes earlier, falling back to a 2-minute IEX median.
• Dashboard:
  • Today's watch, sorted live by confidence.
  • A world-exchanges board.
  • The Trade desk and News ingestion cards.
  • Times in the operator's timezone (CET).
  • Bought/sold amounts and P&L on fills.
• PPO policy (rl/), research harness (research/) and few-shot forecasting harness (fewshot/). The policy learns in shadow and retrains after each close; it trades only after passing its promotion gate.

Changed

• US equities only; crypto was removed.
• Positions: the exit policy is stop, target, strategy exit, end of day, plus the +1R scale-out. There are no adds.
• Watchlist: it starts empty, and no stock list is hard-coded.
• Risk dial: it is saved and survives restarts (default 4 until set). It drives the desk's margin and the agents' stance.
• Deployment: the backend container runs on the host network so it can reach Ollama, and the frontend's nginx proxies through host.docker.internal.

Fixed

• Stops at the broker. The stop and target legs never reached the broker (order_class was missing).
• Manager crash. A watcher read with no price crashed every portfolio-manager cycle, so nothing could trade.
• Missing prices. Picks added before the stream connected never received prices.
• Spread gates read single IEX quotes (BE 7.4% against a real 0.07%) and a 1% placeholder.
• Once per day. The one-entry-per-day rule was lost on restart.
• Reloaded fills now get FIFO P&L.
• Desk calibration. The desk rejected nearly every trade: its prompts leaned toward no and its bar was 0.55. The analyst also anchored on the breakeven, and the critic echoed the analyst's number.

Status

• No edge found yet. The performance component alone shows none over two years (t ≈ −0.7). News, discussion and desk decisions are judged from their logs going forward.
• First live paper day: 4 closed trades, −$96. The flat 0.8% minimum stop sits inside normal minute noise for volatile stocks; volatility-scaled stops are the next change to consider.
• Backtester: it does not model the scale-out yet.
• Requirements: Ollama with qwen3.5:9b and qwen3:4b; without them, entries are blocked.

Full changelog: v0.0.0...v0.1.0 · PR #3

v0.0.0: initial pre-release

First tagged baseline of TradeFlow. Experimental: use at your own risk. This is a side project, not a battle-tested trading tool. Paper trade first.

Highlights

• Quant strategy library (15 strategies), regime detection, adaptive per-moment strategy selection, and a quant council consulted by the manager and the trade bots
• Off-process analytics: Monte Carlo TP-before-stop odds, pairs cointegration, portfolio VaR/CVaR, and ledger ratios
• Latency tracking on the dashboard; tick-to-decision is about 20 µs
• Unfilled-order protection (no stacked buys; stale orders cancelled)
• Classic / stair capital mode (reserve untouched, profit banked per step)
• MIT license

The version stays on 0.0.x until a significant update lands.