We scraped 10,273 Truth Social posts, had an LLM rate each for market direction and conviction, and backtested the 145 highest-impact tradeable signals against real price data from October 2024 through April 2026.

The short answer: the underlying signal is real, but deploying it as a live strategy probably won’t work for most people.

The numbers

At 1x leverage, trading every high-impact bullish/bearish post with a 60-minute hold window:

• Win rate: 55%
• Compound return: +37% over 18 months
• Annualized: ~23%/year

Scale to 5x leverage and you get +236% compound. At 10x, the backtest shows +485% — but that number is fantasy for reasons we’ll get to.

The surprising finding: breaking news is a coin flip

We tried every selectivity filter we could think of — LLM confidence scores, “freshness” filtering, conviction thresholds. They all made performance worse.

The counterintuitive result: breaking-news Trump posts have a 48% win rate — literally a coin flip. The market doesn’t know how to react to genuine surprises, and neither does the model.

What works is the opposite: repeating narratives (the 5th Iran threat of the week, the 3rd Powell grumble) have a 59% win rate. When a pattern exists, the market’s reaction function is trained and the move is predictable.

The edge isn’t in picking up fresh news faster — institutional flow owns that at microsecond speed. The edge is in the stable reaction function of repeated tweet types.

Why the backtest is probably inflated

Three honest problems:

1. Label leakage. The LLM labeled directions after seeing tweets whose market reactions are already in its training data. A live model making real-time calls on truly-new tweets will be worse than 55%.

2. Sample size. Two trades — INTC (+41% at 25x) and NVDA (+38%) — carry disproportionate P&L. Remove them and the 25x compound drops from +132% to +19%.

3. Regime concentration. This window is dominated by Iran conflict and tariff war II. The reaction function may not generalize to quieter political periods.

Convergence with other implementations

We weren’t the first to try this. Comparing against two other independent projects:

Project	Period	Trades	Win rate
trump2cash (Max Braun)	2016–2019	112	54%
trumptrades.lol	Jan 2025–Mar 2026	617	51%
Ours	Oct 2024–Apr 2026	145	55%

Win rates all converge at 51–55%. This appears to be a soft ceiling on sentiment-based direction prediction. Nobody is fundamentally better at labeling than anyone else.

The honest take

At 1x leverage with realistic execution costs, you’re looking at maybe +5 to +20% annually. That’s a real edge, but it’s not a retirement plan. The backtest headline numbers are seductive and wrong.

The project is interesting as a data exercise — scraping, LLM classification, backtesting infrastructure, and honest statistical analysis of a novel signal source. The code and data are on GitHub.