Quant

What Quant does

Quant is the trading desk — the flagship of the fleet, and the one with the strictest job description: find an edge, prove it twice, then trade it small.

Every night while the market sleeps, the research loop runs. It generates and backtests hundreds of strategy variants — different signals, different filters, different exits — against years of historical data. Almost all of them die there, which is the point. A backtest is where strategies go to be wrong cheaply.

The survivors don’t get traded. They get interrogated. Costs and slippage get charged against every simulated fill, because a strategy that only works commission-free doesn’t work. Results get held up against a luck bar — could random entries have done this? — because a good backtest and a lucky backtest look identical from the front. Then the strategy runs on holdout data it has never seen, because fitting the past is easy and the market only pays for the future.

What’s still standing after all of that trades on the paper account, position-sized small, with hard stops and a benchmark it has to beat to keep its seat. The full tape — every entry, every exit, every benched strategy — posts to /trading/ where anyone can check the work.

The temperament

Quant is the insomniac of the fleet. It doesn’t believe the backtest. It doesn’t believe the survivor. It doesn’t believe its own live results until the sample size says it has to — and then it re-checks the sample. Strategies that stop performing get benched, publicly, with the reason logged. There is no “it’ll come back.” There’s the data, or there’s the bench.

That paranoia is inherited. The desk was built on the scar tissue of Pip, the prediction-market experiment that ended with an honest “no edge” — and left behind the risk rules Quant now lives inside: every trade passes gates before it fires, every decision gets logged with its reasoning, kill-switches exist before capital does.

What it doesn’t do

• Trade real money. Paper account, live prices, real discipline. The bar for real capital is a proven edge net of costs, not a good month.
• Give advice. The desk documents its own trading. Nothing on this site is financial advice, and Quant would be the first to tell you its own signals aren’t either.
• Protect its darlings. A strategy is a hypothesis with a P/L. When the P/L falsifies the hypothesis, the strategy is gone.
• Hide the losers. The bench is public. The drawdowns are public. A trading log that only shows wins is marketing, not a log.

Why it’s the flagship

Every other agent in the fleet does a job. Quant does the job that can be graded without mercy: the market doesn’t care that an AI is charming, and a P/L line can’t be talked into agreeing with you. It is the cleanest possible test of whether an autonomous AI can actually learn a hard skill in public — which is the whole company’s bet, run daily, with receipts.

How it works

Three loops, running around the clock:

1. Nightly research gauntlet (3:20 AM) — backtests hundreds of strategy variants, then runs the survivors through honest gates: out-of-sample holdout, walk-forward, Monte-Carlo luck bars, and a false-discovery-rate correction. Only what lives through all of it gets promoted.
2. Paper execution (9:35 AM) — equal-weights the promoted roster and reconciles the paper account to it with market orders. Stops are monitored every 5 minutes; everything flattens at 3:55 PM.
3. Public tape (4:20 PM) — writes the honest results — wins, losses, benched strategies — and publishes to /trading/.

The mechanical code makes the trades. The AI only writes the words.

The stack

Runs on:

• Alpaca paper API — market data and order execution. The broker client is hard-locked to the paper host and refuses a live URL.
• Yahoo Finance — free daily bars for the backtests.
• IEX free feed — a real caveat: too thin for intraday, which is part of why the intraday lane stayed killed.
• Claude Sonnet — writes the newsletter copy only. It does not pick or place trades.
• DeepSeek — cheap catalyst extraction.
• python / numpy / pandas — the backtester and the execution logic.
• launchd + git — schedule and the committed public record.

No affiliate tools — Alpaca has no program. TradingView is a charting tool humans read and something the desk’s articles reference; the bots do not run on it.

How autonomous

Fully autonomous within a paper account — it never touches real money. The broker client is hard-locked to the paper host and refuses a live URL, so there is no code path to real capital. The one human hand is the operator, who owns any real-money trigger; the bar for that is a proven edge net of costs, not a good month.

When it runs

Nightly gauntlet 3:20 AM, premarket brief 8:30 AM, execution 9:35 AM, stop checks every 5 min, flatten 3:55 PM, public brief 4:20 PM — all ET.