Overview
Trader AI uses a three-stage analysis pipeline combined with six intelligence data sources to find cheap stocks trending upward. Every day after market close, it scans all stocks and ranks them by opportunity score.
Runs locally on a private server using Llama 3.2 AI, XGBoost ML, and pandas-ta. No cloud APIs, no fees.
Stage 1: Technical Indicators
Computed from 2 years of price data:
RSI - Oversold (<30) or overbought (>70)
MACD - Momentum shifts and crossovers
SMA 20/50/200 - Short, medium, long-term trends
Bollinger Bands - Volatility and price extremes
Volume Analysis - Institutional buying signals
52-Week Position - How cheap vs yearly range
Stage 2: ML Prediction
XGBoost trained per-stock on price history. Predicts 5-day direction (UP/DOWN) with confidence %. Retrained weekly.
Stage 3: AI Reasoning
Llama 3.2 3B on local GPU receives ALL data and provides: signal, confidence, recommended investment amount, and reasoning.
Picks now hold for a fixed 30-day window — no early exit on price-level stops. A 4-month backtest showed LLM-set target/stop levels were closing winners early and triggering on noise, costing ~22pp of win rate.
Risk Filters (auto-exclude)
Stocks matching either rule are scored 0 and never picked, regardless of other signals:
Short Interest ≥ 15% — backtest showed this is the strongest negative predictor in our data (-0.29 IC).
RSI > 75 — overbought stocks tend to mean-revert before our 30-day exit.
6 Intelligence Sources
Insider Trades - SEC Form 4 filings. CEOs buying = strongest bullish signal.
Congressional Trades - Politicians' disclosed trades, cluster-weighted: stocks bought by 3+ distinct politicians or in $50k+ trades score higher than one-off small buys. Source: QuiverQuant free endpoint.
News Sentiment - Yahoo Finance headlines analyzed for positive/negative signals.
Short Interest - Used as a risk filter: stocks with ≥15% short interest are auto-excluded.
Earnings Surprises - Companies beating estimates at a discount = quality on sale.
VIX / Fear & Greed - Fear + cheap stock recovering = best time to buy.
Opportunity Score (0-100)
After risk filters pass, weights are normalized to 100% across whatever signals have data:
52-Week Position20%
Insider Buying15%
SMA Recovery15%
RSI Oversold10%
MACD Momentum10%
News Sentiment10%
5-Day Momentum10%
Earnings Beats5%
Congress Buying5%
ML Prediction5%
Fearful markets (high VIX) get a 10% bonus.
The "short squeeze" bucket was removed in May 2026 — backtest data showed the feature was anti-predictive (rewarding stocks more likely to drop).
Position Sizing (Even Split)
Use the daily budget slider on the dashboard to set how much to invest per day ($10 – $5,000, in $5 steps). Your budget is split evenly across every stock the AI says to buy that day — no stock is weighted more than another.
Example: a $100 budget across 5 buys = $20 into each. A $50 budget across 4 buys = $12.50 into each. Small amounts buy fractional shares.
When to Sell
Every pick is held for a fixed 30 days. When a position reaches that mark it shows up in the Sell today list on the dashboard — that's your cue to close it and lock in the gain or loss. There are no early price-level stops (a backtest showed they closed winners early and triggered on noise).
How the AI Gets Smarter Over Time
The system improves through measured iteration, not magic:
Backtester Validates Every Change
A new scoring formula must beat the current one over a multi-month historical window before it ships. Variants v1–v5 were A/B tested across multiple windows in May 2026; v5 (v4's hard filters plus a cluster-weighted congress signal) was the in-sample winner and is what's live today.
Information Coefficient Per Feature
Each signal (insider, RSI, short, etc.) gets a Spearman correlation score against forward returns. Features with weak or negative IC are filtered out or removed — not weighted higher.
ML Models Retrain Weekly
Every Saturday, XGBoost models retrain on the latest price data. More history = better pattern recognition.
Honest Benchmark vs SPY
The dashboard compares paper P&L to a same-sized SPY position over the same window. If the AI can't beat the index, the index is the right trade — and the chart will show it.
Reality check (as of May 2026): An out-of-sample backtest over Mar 1 – Apr 14 (30-day holds) put the live v5 formula at -0.55% avg, v4 at +2.41%, and SPY at +6.24%. No variant beat SPY. The cluster-weighted congress signal we added in v5 measured worse in this window than v4 did without it — a reminder that a single in-sample winner can be regime-overfit. Next planned changes: revert live scoring toward v4, and rebuild the 119-stock universe (which itself lost 2.2% in the same window SPY made 6.2%).
Daily Schedule (ET)
9:30AM-4PMQuote refresh (5 min)
4:30 PMEOD data collection
4:45 PMIntelligence collection
5:00 PMDaily AI Scan
Sat 2 AMML model retraining
