Guessing stocks is noisy and rarely explains itself. Ashwin Paranjape, founding AI lead at Samaya AI, told AI Engineer his team sidestepped that trap by testing agents on earnings prediction, where reported revenue provides quantitative ground truth and the earnings call explains why a beat or miss happened.
The target is consensus estimate, the pooled average of Wall Street forecasts published before earnings.
The test harness is a point-in-time gate that is a programmatic control which cuts data at a fixed minute about five sessions before earnings and blocks future information with an authentication token, Paranjape said. Inside that gate the agent gets company documents, prices, consensus and timestamped news streams, because open web search cannot be gated cleanly, plus expert guidance on how to use context for forecasting.
Across more than 450 earnings runs from July onward, a cutoff Paranjape tied to Fable 5.1 with June knowledge cutoff, consensus error was 1.5 times historical volatility and 1.18 after correcting for the tendency of companies to beat, Paranjape said. He said GPT-6 Astra beat even that bias-corrected consensus on revenue, the metric with the fewest accounting adjustments, and led on hit rate and on surprise correlation, a combined measure of direction and magnitude of the predicted beat or miss.
Macom, a semiconductor supplier to data centers, showed the win case when the agent found a small UK wafer supplier signal, walked it down to the P&L and forecast well above consensus and close to actual, Paranjape said. Ross Stores showed the failure mode when the agent spotted a peer tariff refund but left it out of gross margin and hugged consensus while actual printed higher.
Live data mattered most, as parametric memory alone failed despite June training that should have memorized prior filings, and data frozen 11 weeks before earnings scored 12 to 15 points lower, a gap Paranjape equated to the jump from Sonnet to Fable 5.1. With consensus redacted from retrieval and documents, he said Astra held its accuracy while older models fell, which suggests frontier models can reconstruct the Street view rather than copy it.
That lab test sits beside FrontierFinance, which per the Fireworks Specialized Intelligence Index covers 220 expert-written investment research questions across 12 models from screening to earnings. Watch the alpha product test Paranjape described, especially whether customer internal data and learning from past misses extend the edge from earnings to guidance.
