Strategy research, automated
Same job, different bosses. You can hand it a RSI idea and ask it to backtest. You can put Claude on a cron. Or you can try the hard version: a full-time researcher that invents its own ideas and never waits for you.
A standing prompt on a timer. “Read what moved. Propose one testable idea. Backtest. File a note.” No desk. No swarm. Just the usual man, on a schedule.
A model with tools, a data feed, a sandbox, and a place the note lands (Telegram, Notion, a repo). Hermes / Claude / cron is enough.
The prompt is vague, the backtest is sloppy, or nobody reads the morning dump. Cheap to run. Easy to ignore.
You bring the hunch: “these indicators, this regime.” The agent codes it, backtests, tweaks parameters, writes variants, keeps a graveyard of what failed.
A clean research notebook pattern. Walk-forward rules. A critic that is not the same voice that proposed the trade.
It overfits the seed. Ten “improvements” that all peek at the same window. Useful junior. Not a partner yet.
The agent sits where you sit. Clicks charts, opens a notebook, pulls a series, writes a cell, stares at a residual. Research as computer use, not as a chat.
Browser or notebook control, screenshots or DOM, and a hard rule: it may look, it may code, it may not silently place size.
The UI shifts, the session dies, or it narrates charts instead of producing a test. Looks magic. Often slow.
It hunts old and new papers so you don’t need homies for literature. Read, extract the claim, code a cheap replica, backtest on your book, file keep / kill / later.
A paper feed, a reproduction budget (hours, not weeks), and a house style for “what counts as the same idea.”
It implements abstracts, not methods. Or it treats every PDF as alpha. Most papers do not survive costs and a new decade.
Not one brain. An idea person, a data person, a backtester, a risk officer, a writer. Separate prompts, separate context, instrumented handoffs. The “digital quant” people mean.
Role contracts. Shared artifacts (hypothesis card, code, tear sheet). A human or a gate that can say no. See also RD-Agent-style research + build split.
Everyone is the same model agreeing with itself. Or the critic sees the in-sample Sharpe before it “independently” reviews.
Keep a zoo of strategies. Mutate the good ones. Protect weird ones that are merely different. Quality-diversity, not “one Sharpe to rule them all.”
A cheap evaluator, a diversity metric (regime, holding period, factor family), and disk for the graveyard. LLMs write the mutants; evolution keeps score.
The fitness function is a leaky backtest. Then you evolve a poet of in-sample luck.
Mass-produce signals, not stories. Generate expressions, score IC / decay / crowding, regularize so you don’t rediscover last year’s popular factor with extra parentheses.
A factor language, a point-in-time panel, turnover and capacity checks. Human-in-the-loop versions exist (Alpha-GPT). Fully closed loops exist too.
The factory ignores capacity. A beautiful IC that cannot take money is a museum piece.
Agents compete. One proposes, one tries to break it: leakage, lookahead, regime luck, “it only works on Tuesdays in 2017.” Only survivors get a writeup.
A hostile critic with different tools. Explicit kill criteria. Reward for finding the cheat, not for being nice.
The red team is polite. Or both sides optimize the same scoreboard and call it science.
Backtests propose. A paper book decides. Nothing is “done” until it lives through new days with costs, slips, and the ugly fills the simulator forgot.
A paper account, a promotion rule, and the spine to wait. Most agent papers skip transaction costs and universe handling. Don’t.
You promote on a two-week hot streak. Live is just another small sample if you are impatient.
The brief you actually want. Completely hands-off. It is not given a thesis. It has to find one: papers, anomalies, regime shifts, failed ideas worth a second look. Then it runs 04–09 on itself.
A world model of what “good” means for this book. Memory so it does not rediscover last month’s corpse. A budget. A kill switch. Someone still owns risk.
Autonomy without a scoreboard becomes a blog. Or the scoreboard is Sharpe and it becomes a p-hacking engine with a nice voice.