The Null Result

Published: October 2025

Most startups in the AI alternative data space do not publish their failures. They highlight the backtests that worked and quietly bury the models that blew up in live trading.

We are publishing ours.

The Original Thesis

In our V1 architecture, we believed that fine-tuning a Large Language Model (LLM) on historical earnings calls would allow it to reliably predict post-earnings drift (PEAD). The backtests were incredible. The Sharpe ratio was 3.1.

The Failure

In live trading, the model lost money. Not a little money. It underperformed a basic S&P 500 index fund by 12% over three months.

Why? Because the model hallucinated confidence. It was optimizing for text completion, not for structural market reality. When a CEO used optimistic language about a fundamentally broken supply chain, the NLP model bought the optimism. It lacked a deterministic firewall.

The Rebuild

We threw out the V1 architecture entirely. We stopped trying to predict the market using end-to-end black box models.

Instead, we built Candlery V2: a deterministic extraction engine. We now use NLP onlyto parse unstructured text into structured JSON intents. We then pass those intents through a hard-coded, quantitative risk firewall. If a signal does not have a mathematically verifiable historical analog, it is rejected.

We prefer false negatives over false positives. We prefer missed opportunities over capital destruction.