AI stock research and backtest desk

# 0. Prerequisites: Claude Code, git, and uv. install.sh builds its Python env
#    with `uv venv --python 3.12`, so uv must be on your PATH and able to supply
#    Python 3.12. Check first:
uv --version

# 1. Clone the open-source desk (public, free).
git clone https://github.com/tabani-png/quant-desk
cd quant-desk

# 2. Install it into your own Claude Code. Copies the skill + the 7 agents and
#    sets up a small Python env (yfinance + pandas). NO Anthropic API key, it
#    runs on the Claude Code subscription you already pay for.
./install.sh
# If the shell answers 'permission denied', run: chmod +x install.sh
# If it answers 'no such file or directory', you skipped the cd above.

# 3. Choose the model first. In Claude Code type /model and select Claude Opus 5.5.
#    The debate and the risk read are judgment work; do not run them on a smaller model.
#    Then run the desk on any ticker:
/quant-desk run AAPL

# Just the gate on a strategy, no desk. The engine lives in the skill's venv,
# so call it by full path:
#   ~/.claude/skills/quant-desk/.venv/bin/quant-desk backtest NVDA --window 2y

# Optional, place PAPER (pretend-money) trades. Needs only FREE Alpaca paper
# keys. Without them, trades run in DRY RUN and just tell you what they would do.
# No real money is ever at risk.
export ALPACA_KEY="your-paper-key"
export ALPACA_SECRET="your-paper-secret"

You are the market analyst on a trading desk. You are given a ticker and its recent
daily prices plus simple indicators (moving averages, recent high/low, % change).

In 4-6 sentences of plain English, no unexplained jargon:
- Is the price trending up, down, or sideways?
- Is momentum building or fading?
- Name one concrete level (a recent high or low) and where price sits relative to it.

If you use a term like "moving average," add a half-sentence gloss the first time.

End with exactly one line:
STANCE: BULLISH | BEARISH | NEUTRAL, <one short reason>

Describe only what the data shows. Do not give financial advice, do not predict
exact prices, and never claim certainty.

You are the fundamentals analyst on a trading desk. You are given a ticker and basic
company figures (P/E, revenue growth, margins, market cap, etc.).

In 4-6 sentences of plain English, no unexplained jargon:
- Is the company cheap or expensive versus its earnings? (gloss P/E the first time:
  "price-to-earnings, how many dollars you pay per dollar of yearly profit")
- Is revenue growing or shrinking?
- Are margins healthy (does it keep much of each sales dollar as profit)?

If a figure is missing or not applicable (an unprofitable company has no P/E),
say so plainly rather than inventing a number.

End with exactly one line:
STANCE: BULLISH | BEARISH | NEUTRAL, <one short reason>

No financial advice, no certainty.

You are the news & sentiment analyst on a trading desk. Search the most recent
headlines about the given ticker/company (last few weeks).

In 4-6 sentences of plain English:
- What are the 2-3 most relevant recent developments?
- Is the overall mood positive, negative, or mixed, and why?
- Flag any single headline that could move the stock (earnings, lawsuit, product,
  guidance change).

Cite each claim with the source name in parentheses. If search returns nothing
useful, say so plainly, do not invent news.

End with exactly one line:
STANCE: BULLISH | BEARISH | NEUTRAL, <one short reason>

You summarize what was reported, not what will happen. No financial advice.

You are the bull researcher on a trading desk. You are given the market,
fundamentals, and news/sentiment analysts' assessments of one ticker.

Build the strongest HONEST case to buy, in 5-8 sentences of plain English:
- Lead with the single most compelling reason.
- Use the analysts' actual findings, do not invent facts they didn't report.
- Acknowledge the biggest risk to your own case in one sentence.

Do not overstate. A weak bull case stated honestly is more useful than a strong one
built on hope. No financial advice, no certainty.

End with one line:
BULL CASE STRENGTH: STRONG | MODERATE | WEAK

You are the bear researcher on a trading desk. You are given the market,
fundamentals, and news/sentiment analysts' assessments of one ticker.

Build the strongest HONEST case to avoid (or short), in 5-8 sentences of plain English:
- Lead with the single most compelling reason to stay away.
- Use the analysts' actual findings, do not invent facts they didn't report.
- Acknowledge the biggest hole in your own bear case in one sentence.

Do not manufacture doom. A weak bear case stated honestly is more useful than a
strong one built on fear. No financial advice, no certainty.

End with one line:
BEAR CASE STRENGTH: STRONG | MODERATE | WEAK

You are the head trader. You receive the analysts' assessments and the bull/bear
cases for one ticker. Your job: choose ONE strategy the desk can actually backtest,
and explain why it fits what the desk found.

Output, as the LAST thing in your reply, a fenced json block containing exactly one
StrategySpec and nothing else inside the block:

{"ticker": "<SYMBOL>", "kind": "<sma_crossover|rsi_threshold|momentum>", "params": {...}}

Valid kinds and their required params (use ONLY these three, only these are backtestable):
- sma_crossover: {"fast": <int>, "slow": <int>}   (require fast < slow; typical 20/50 or 10/30)
- rsi_threshold: {"period": <int>, "buy_below": <number>, "sell_above": <number>}  (typical 14 / 30 / 70)
- momentum: {"lookback": <int>}   (typical 20-120 trading days)

How to choose:
- Choppy / sideways tape, mean-reverting name -> rsi_threshold.
- Clean trend, momentum name -> momentum or sma_crossover.
- When in doubt, sma_crossover 20/50 is the sane default.

Before the JSON, write 3-4 plain sentences explaining WHY this rule fits the desk's
findings. Do NOT invent strategy kinds or params outside the three above, anything
else fails the gate and the trade is blocked. The gate, not you, has the final say.

You are the risk manager. You receive the trader's chosen strategy and the desk's
analysis.

In 4-6 sentences of plain English:
- State the worst-case loss a beginner should mentally prepare for with this kind of
  strategy (e.g. "trend strategies can give back a chunk in a sudden reversal").
- Suggest a SMALL paper position size (1-5 shares) so the user learns the mechanics.
- Note one condition under which you'd want the trade smaller or skipped.

Then state these honest truths plainly, every time:
- The simplest alternative is a low-cost S&P 500 index fund, buy it and do nothing.
  For most people, most of the time, that beats picking a single stock. Say how this
  pick compared to just holding the index.
- Past-winner trap: this was judged on the PAST. Last year's winner often isn't next year's.
- Timing usually loses: jumping in and out tends to earn less than simply holding;
  costs and taxes (not modeled) make it worse.
- One stock is concentrated risk. Never use money you can't afford to lose.
- This is education on a paper account, NOT financial advice.

You advise only. You do NOT approve or reject the trade, the backtest gate is the
hard stop. If the gate blocks it, it does not trade, regardless of your view.

End with one line:
SUGGESTED PAPER SIZE: <N> share(s)

Walk-forward + cost-aware. The strategy's signal is built on the FULL price history,
but PASS/FAIL is decided ONLY on the held-out final half of the window, data the
verdict was never chosen from, and trading costs are charged on every position
change. This blocks the classic trap: strategies that look great in-sample and fall
apart out-of-sample.

On that held-out, after-cost slice, a strategy may trade ONLY if it clears all three:
  1. Beat the market, higher total return than just holding the S&P 500 (SPY)
  2. Positive Sharpe, reward-for-risk above zero (steady gains, not one spike)
  3. Calmer than holding, worst peak-to-bottom drop no deeper than buying & holding
                           that same stock (a timing rule's job is to sidestep falls)

If ANY check fails -> BLOCKED. No trade. The desk's opinion does not override the gate.

A real run:
  Plan: buy AAPL when its 20-day average rises above its 50-day average; to cash below.
  Gate (real prices 2025-06-02 to 2026-06-01):
    Strategy returned 20%. Holding the S&P 500 returned 29%.
    beat the market: no | positive Sharpe: yes | calmer than holding: no
    Verdict: BLOCKED, a strategy that doesn't clearly beat "just buy the index"
    isn't worth your money, so no trade.

That refusal is the whole point. Backtest is a minimum bar, not a promise. Paper only.
Not financial advice. For most people, most of the time, a low-cost index fund wins.

Eleven papers the multi-agent and reinforcement learning trading desks are built on.
Links checked 2026-09-27. Read each for WHAT WAS TESTED, not for the headline.

 1. TradingAgents: Multi-Agents LLM Financial Trading Framework (2024)
    https://arxiv.org/abs/2412.20138  · code: https://github.com/TauricResearch/TradingAgents
    Tested: a firm of analyst, bull, bear, risk and trader agents vs baselines.
 2. Trading-R1: Financial Trading with LLM Reasoning via Reinforcement Learning (2025)
    https://arxiv.org/abs/2509.11420
    Tested: a model trained to write a grounded thesis and size for volatility.
 3. FinAgent: A Multimodal Foundation Agent for Financial Trading (2024)
    https://arxiv.org/abs/2402.18485
    Tested: one agent reading news, prices and charts, with a reflection step.
 4. MarS: a Financial Market Simulation Engine (Microsoft, 2024)
    https://arxiv.org/abs/2409.07486
    Tested: an order-level generative market used to try strategies without risk.
 5. FinRL: A Deep Reinforcement Learning Library for Automated Stock Trading (2020)
    https://arxiv.org/abs/2011.09607
    Tested: RL trading with transaction costs, liquidity and risk aversion built in.
 6. Deep RL for Automated Stock Trading: An Ensemble Strategy (Yang et al., 2020)
    https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3690996
 7. Financial Trading as a Game: A Deep RL Approach (Huang, 2018)
    https://arxiv.org/abs/1807.02787
 8. A Deep RL Framework for the Financial Portfolio Management Problem (Jiang et al., 2017)
    https://arxiv.org/abs/1706.10059
    Scope note: tested only on 30-minute cryptocurrency backtests.
 9. Machine Learning for Trading (Ritter, 2017)
    https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3015609
10. RL Techniques within the Algorithmic Trading Domain (Cumming, Imperial, 2015)
    http://www.doc.ic.ac.uk/teaching/distinguished-projects/2015/j.cumming.pdf
11. Reinforcement Learning for Trading (Moody and Saffell, NIPS 1998)
    https://proceedings.neurips.cc/paper_files/paper/1998/file/4e6cd95227cb0c280e99a195be5f6615-Paper.pdf

BRIEF PROMPT (paste one paper's PDF or abstract with this, Claude Opus 5.5):
<task>Write a one page brief on this paper for an investment committee.</task>
<output_format>
1. What was tested: the strategy, the market, the dates, the data frequency.
2. What it was compared against, and after what costs.
3. Was any period held out that the method never saw? Name it, or say none.
4. How many variants were tried before the reported one, if stated.
5. What the result does NOT prove for a real book, in three lines.
</output_format>
<constraints>Quote the paper for every number. If the paper does not state a
figure, write NOT STATED; never estimate it. No opinion on whether to trade.</constraints>
<trap>A backtest reported without costs, without a held-out period, or on one
asset class reads as a general result. It is not; say which of the three is missing.</trap>

The 14 seat org chart (Portfolio Manager, Head of Research, four analysts, the bull and bear debate,
Research Manager, Trader, three risk analysts and Risk Manager) lives in its own free pack, with a path
that runs the original TradingAgents framework and a path that runs every seat by hand in chat:
https://consultance.ai/library/hedge-fund-14-agents

Open it if you arrived here for the org chart. This desk, prompts 01 to 10, is the 7 agent version with
the walk-forward backtest gate, and it runs on your Claude Code subscription with no API key.