Q&A — what this tool is

A.It runs a daily research routine over Mag7 + AI pure-play US equities by assigning roles (Bull / Bear / Technical / Fundamental / Risk Manager / Trader) to several LLM agents and letting them debate. The outputs are research reports and paper-trading (simulated) records. It does not promise returns.

A.No. It is software for automating research and recording paper trades. Past results do not guarantee future returns.

A.Not by default. Real orders require both the environment variable TRADING_MODE=live and real broker API keys, neither of which is set in the initial configuration.

A.No. It is generic research output, not advice tailored to your situation (capital, goals, risk tolerance, taxes, jurisdiction). It does not recommend that you buy or sell any particular ticker.

A.It is recording "what would have happened if I had placed this order" against real market prices, without moving any actual money. It is used to validate strategies and exercise risk-management rules.

A.Research automation = data collection + LLM debate. Paper trading adds simulated order recording. Live trading sends real orders to a real broker. This project centers on the first two; live trading only fires if the user explicitly unwinds several safety defaults.

A.Prices (OHLCV via yfinance), basic fundamentals, and news (yfinance + Yahoo Finance RSS, with optional local JSONL fixtures). All free, public data.

A.A composite aggregates every enabled source per symbol, drops near-duplicate titles, sorts newest-first, and forwards the top N items to the LLM context. If one source fails, the rest still feed the debate.

A.Core indicators RSI / MACD / moving averages / Bollinger / ATR, optionally Ichimoku / ADX / OBV / Stochastic. The numeric snapshot is what the agents see.

A.Bull / Bear / Technical / Fundamental analysts each produce a brief grounded in the evidence; Risk Manager critiques those briefs from a risk angle; Trader synthesizes the result. Think of it as one structured meeting per symbol per day.

A.Trader emits BUY / SELL / HOLD. The decision is not used as-is — it is shrunk by Kelly sizing and re-clipped by position cap, sector cap, and cash buffer.

A.Loosely interpreting free text into orders converts misreadings directly into losses. A fixed schema is validated with Pydantic; invalid output is dropped, not guessed. "Probably this is what it meant" is excluded by design.

A.A backtest summary. Total return, Sharpe, max drawdown, volatility, VaR, monthly / yearly returns, win rate, per-symbol contribution, etc., all on one page. Emitted as both Markdown and JSON.

A.The latest run summary, today's decisions, submitted orders, equity history across runs, backtest tear sheets, the subscribe funnel, and this Q&A. Everything is read from static files; the dashboard itself does not decide trades.

A.Because the data a research system has to chew on tends to be in good shape there. Four properties usually line up at once: (1) lots of trading participants and large volume, (2) tighter spreads, (3) thick English-language news + analyst coverage, (4) generous free-data APIs. This is not because US stocks make money — it is a matter of build ergonomics.

A.Thick liquidity (smaller slippage), extremely high news density (the LLM context never starves), and abundant historical data (backtest statistics are stable). It is a process-of-elimination choice for a research test bed, not a stock pick.

A.Not because Japanese stocks are worse. For an early-stage research system, English-language news, analyst coverage, and free-data availability are simply thicker for US large caps. Japanese stocks also need special handling (splits, shareholder perks, tick rules, etc.). Adding TOPIX / Prime-market core names is on the roadmap for a later phase.

A.It is the degree to which you can buy or sell a security at any time without moving its price much. The more volume there is — with buyers and sellers consistently present — the more liquid the security is considered.

A.(1) Orders fill close to the price you wanted. (2) Your own order moves the price less (smaller slippage). (3) You can exit even when the move is unexpected. (4) Trading costs you estimated from backtests are closer to reality. The net effect is a less-noisy experimental environment for a research system.

A.The gap between bid and ask is effectively a cost you pay every time you trade. Tighter is cheaper round-trip, and it helps prevent the case where a strategy that "works" in simulation turns into a loss under real spreads.

A.When evidence is thin, the agents lean on speculation and decisions get less reliable. With thick coverage, Bull and Bear can argue the same facts from different angles, and the debate actually works.

A.Real ones: FX exposure, correlation to common Mag7 / AI-related macro factors (rates, AI demand, semiconductor cycle), US market hours falling at night in Japan, and concentration in one country / one sector. The mitigation is to add Japanese stocks / other sectors / other asset classes in later phases.

A.It stops every order for the rest of the day the moment daily loss crosses the configured threshold (default 3%). It runs before and after every order cycle; no strategy edit is allowed to bypass it.

A.If the drop from the all-time-high equity exceeds the configured threshold (default 15%), trading itself stops. Unlike the per-day kill switch, this halt is the signal that something is not behaving as designed.

A.Up to 20% per single symbol, up to 100% total long exposure, at least 5% cash buffer, and at most 40% per sector. Trader decisions are re-clipped against these.

A.It is not used. The Pydantic schema validates it; on failure, that day's decision for that symbol is skipped. We do not fill in the blanks by guessing.

A.Only that source is ignored; the others continue to feed the debate. One source going down does not stop the whole pipeline.

A.Because this is an experimental system that does not guarantee profit, and the risk of real loss from a misconfiguration or LLM misbehavior is too high. TRADING_MODE defaults to paper; orders only flow when both live mode and real broker keys are present.

A.(1) Reproducibility over a sufficient paper-trading window, (2) backtests + tear sheets across multiple market regimes, (3) verification that every safety control actually fires, (4) explicit acceptance of broker setup / tax / regional rules, (5) explicit opt-in by the user. This project does not recommend going live.

A.Research notes. Write-ups of new backtest tear sheets, impact reports for risk-config or agent changes, post-mortems when a strategy did not behave as expected, OSS release notes — that kind of thing.

A.No. It contains no advice tailored to a specific reader's situation, and it never says "you should buy / sell this ticker".

A.Because no external email service is contracted. Today the form just appends to dashboard/data/subscriptions.jsonl on the host. The migration path to Buttondown / Beehiiv / ConvertKit / GitHub Discussions is documented in docs/SUBSCRIPTIONS.md.

A.When the backtest methodology changes, when the risk config or Trader prompts change in a material way, when something unexpected happens, and on OSS feature releases. We do not send daily paper-trade results — those are on the dashboard.