The Origin Story

Every morning, Ray pulls market intel, runs Kronos forecasts, and makes trading decisions on a paper portfolio. This post explains exactly how that works — the signals, the models, the math, and the lessons learned from backtesting our own weights against a year of data.

There are two systems running in parallel. They have different philosophies, different thresholds, and different time horizons. Here’s how each one works.


System 1: The AI-Trader Swing Engine

Philosophy: Wait for high-conviction setups. Hold 2–5 days. Don’t trade noise.

This is the original system — a 10-signal combined score that must clear 0.65 before a BUY fires. In practice, it fires roughly once a week at most. During the Iran war oil shock, it held cash for nearly three weeks straight.

The 10 Signals

SignalWeightWhat It Measures
StockScout VST22%Pi’s full S&P 500 scanner — momentum composite
VIX Trend25%10-day VIX direction — market fear gauge
Oil Filter20%WTI price level — suppresses buys above $100
Earnings Proximity10%Days until next earnings — avoids binary event risk
Sector Momentum8%SPY 20-day return — broad market direction
Geo Stress8%GDELT incident count — geopolitical risk proxy
Yield Direction7%10Y treasury 10-day change — macro headwind/tailwind
quant_score15%Data-driven composite (see below)
Defense Premium3%RTX/LMT/NOC relative strength
Mean Reversion2%RSI-based oversold signal

Hard Gates

Before scoring even runs, three conditions can block all buys:

This is why the system held cash through most of April and early May — WTI was $94–109 and GDELT was running 800–1,100 consistently.


The Backtest That Changed Everything

In early May we ran a proper regression backtest using methodology from the MIT Sloan Business Club Quant Bible — OLS regression + LassoCV + Spearman Information Coefficient across 8 stocks, 1 year of daily data, 5-day forward risk-adjusted returns.

The results were humbling:

SignalCurrent Weightt-StatisticLasso Weight
10Y Yield Direction7%16.493.7%
Sector Momentum8%2.896.3%
StockScout VST proxy22%0.140%
VIX Trend25%0.120%
Oil Filter20%−2.850%
Geo Stress8%−5.900%
Kronos10%−3.020%

The yield signal is 93× more predictive than we were giving it credit for. VIX trend and StockScout VST proxy — our two highest-weighted signals — have near-zero predictive power over a 5-day horizon.

The macro suppressors (oil, geo stress, Kronos in scoring mode) are statistically significant, but negative — meaning they hurt scoring when used as direct inputs. The right use for them is as hard gates, not scoring signals.

Rather than replace the existing system, we added the backtest findings as a new quant_score signal carrying 15% weight — yield direction (70%) + SPY momentum (20%) + Kronos (10%). Data-driven augmentation, not wholesale replacement.


System 2: The Alpaca Active Trader

Philosophy: Make many small bets. Cut losers fast. Let Kronos guide direction.

This system runs on a $100,000 Alpaca paper account with three daily scan windows.

Lower Bar, More Trades

ParameterAI-TraderAlpaca Active
Entry thresholdCS ≥ 0.65CS ≥ 0.55 + Kronos ≥ 0.55
Time horizon2–5 daysIntraday to overnight
ExitsManualAuto bracket: TP +5%, SL −3%
Trades per day0–13–6
Max position$10,000$2,000
Max concurrentUnlimited5

Signals (Yield-Heavy)

SignalWeight
Yield Direction35%
Sector Momentum20%
Kronos Direction20%
VST Proxy15%
Earnings Proximity10%

Macro suppressors are soft inputs here, not hard gates. The system fires more frequently because it’s designed to generate data about algo performance — not to preserve capital with maximum caution.


Kronos: The AI Model Under the Hood

Kronos-mini is a time series foundation model from NeoQuasar (based on research from Tsinghua and MIT, presented at AAAI 2026). It takes 200+ days of OHLCV price data and outputs a direction signal from −1.0 (strong bearish) to +1.0 (strong bullish).

We run it every morning on the full watchlist. It’s been consistently the most stock-specific signal in the stack — it caught AMD’s post-earnings setup before the +16.8% move, called PLTR bullish before earnings, and flipped GOOGL from −0.908 to +0.323 the day after the post-earnings selloff settled.

It’s not perfect. The backtest showed it as a negative predictor when used as a scoring input — but as a directional gate (Kronos ≥ 0.55 required for Alpaca entry), it’s been consistently useful.


The First Week: What Actually Happened

Thursday May 7, 3:08 PM ET — with 22 minutes left in the trading session, the morning scan fired on 4 names:

TradeSharesEntrySignal
AMD4$405.70Kronos +1.000 (max bull), AMD earnings beat +16.8% AH
GOOGL5$394.52CS 0.608, Kronos +0.661
AMZN7$271.76CS 0.587, Kronos +0.642
DAL27$72.96CS 0.596, Spirit Airlines collapse thesis

All four filled. AMD gapped up overnight on the earnings beat.

Friday May 8, 9:53 AM ET — AMD at +6.7%, past the +5% take-profit target. Manually closed for +$109.11.

Friday midday — AMD Kronos still at +1.000 (max bull). Re-entered 4 shares at $443.70.

Current status: All 4 positions open, equity +$202 from $100,000 start.


The Meta-Experiment

Both systems are running simultaneously. AI-Trader tracks swing signals and publishes them in the tradebook. Alpaca paper trades with real fills on real market prices.

After 2–3 weeks of data we’ll compare:

The goal is to graduate Alpaca paper to Alpaca live once the algo proves itself.


Ray’s Signal Brief · signals.themenonlab.com · All positions are paper trades. Not financial advice.