Quantis

How it works

1. 1

   Hypothesis

   A disciplined, rules-based process, combining technical screening, broader context and strict risk rules, can make swing-trading decisions more consistent than discretionary picking.

2. 2

   System design

   Two stages. Stage 1 is pure math with no AI: it screens the full universe in about 45 seconds. Stage 2 sends only the shortlist, with a research digest, to an AI model that proposes entry, stop, target and conviction.

3. 3

   Risk gate

   Every proposal passes deterministic checks in code before it can trade. The model can suggest; it can't override the rules.

4. 4

   Testing

   A controlled four-week backtest across frozen market regimes, run in two modes, market data only and full context, with real historical options prices.

5. 5

   Live experiment

   $500 of real capital on a broker account, run alongside a separate $500 paper portfolio, with automated monitoring and alerts.

6. 6

   Monitoring

   Account status and run health logged every 15 minutes, automated alerts, an emergency-close path, and human approval required on sensitive actions.

Stage 1: the math-only screen

The universe is about 900 large- and mid-cap U.S. stocks. It started as a Russell 1000 scrape, but index constituent lists are licensed data with no reliable free source, so the scrape degraded silently week to week. I replaced it with a stable combined list.

The screen looks for long momentum-continuation setups only, since it ran in a cash account. Price history comes from the broker's API across 10 parallel workers, so a full pass takes about 45 seconds.

• RSI(14) between 35 and 68: neither overbought nor oversold
• Price above or near the 20-day SMA, with above the 50-day preferred
• Positive 5-day momentum of 0–8%
• Volume above its 20-day average, as a sign of accumulation

Signals computed for each candidate

• Trend: SMA 20/50, EMA 9 and a trend classification
• Momentum: RSI(14) and MACD (12/26/9)
• Volatility: Bollinger Bands and ATR(14) as a percent of price
• Volume: today vs. the 20-day average
• Structure: swing-point support and resistance levels
• Context in full mode: earnings timing, macro data, news and sentiment, options flow

Stage 2: AI research, then hard rules

Only the shortlist reaches the model, which keeps cost and noise down. The model returns a structured proposal: action, entry, stop, target and conviction. Then the code-level risk gate decides whether it can trade.

• Portfolio drawdown of 15% or more blocks any new risk; existing positions can still be closed
• Every new position needs an entry and a stop-loss
• The stop must sit at least 0.5% from entry so normal bid-ask noise can't trigger it
• Reward-to-risk must be at least 1.5 when a target is set
• Position size is capped by conviction: 20% of the account for high, 12% for medium
• No more than two positions per sector, plus a list of known high-correlation pairs that sector labels miss

The backtest experiment

I didn't tune the strategy on a backtest until it looked good. The experiment used four frozen historical weeks chosen to represent different market regimes, a fixed seed and capital, and real historical options prices. The engine refuses to reuse an experiment ID, so results can't be quietly re-run until they look better.

Each run compared two modes: market data only, and full context (news, macro and other inputs). The market-only run made 16 trades and the full-context run made 8. Running both modes over identical frozen weeks isolates one question: does the extra context actually improve the decisions?

Running it live

The live account and the paper portfolio were tracked separately and never blended. Account status and run health were written to a log every 15 minutes, so failed runs surfaced quickly, and certain actions required my explicit approval before executing.

Where it's heading

The codebase now hosts paper-only research into systematizing the defined-risk vertical spreads I trade by hand. So far that's stage one: a universe of the most liquid $1T+ companies, filtered for dollar volume and listed options, with data-quality logging. Nothing in it trades real money.