A developer going by "louidev" spent 38 days running a cronjob against Google's Gemini 2.5 Pro, collecting live stock price forecasts and publishing the results on Hugging Face. The dataset, louidev/glassballai, has over 1,140 rows — roughly 30 predictions per trading day across multiple tickers, each covering a 10-day forward window. The author's dataset card is direct about why provenance matters: the forecasts were captured as they happened, against live search-augmented context, and there's no way to reproduce them after the fact.
The pipeline runs <a href="/news/2026-03-15-agentpages-autonomous-research-publisher-gh-aw">two agents in sequence</a>. A factual research agent handles grounded news search, pulling analyst price targets, known catalysts, and macro risk factors — all with mandatory inline citations. That output feeds a financial reasoning engine built around four hard-coded principles: Skeptical Realism, Bilateral Analysis, Temporal Synthesis, and Historical Synthesis. The last principle draws an intentional distinction between grounded claims, which require citations, and the model's internal reasoning about market cycles, which explicitly does not. Every forecast ran at temperature=0. That choice was deliberate: with randomness removed, any drift in predictions or narrative over 38 days reflects something about the model's temporal behavior, not random sampling.
Each row captures the predicted price, a natural-language rationale, a sentiment score between 0.0 and 1.0, and a self-reported confidence float, alongside the full system prompts and model config. Actual closing prices aren't included — licensing constraints — but the author describes them as "rehydratable" from public market data. The dataset ships under CC-BY-NC-4.0, with a companion dashboard at glassballai.com and a Colab quickstart notebook. Whether it holds up as a research resource depends on what questions you're asking: forecast accuracy against actual prices is only one of them, and probably not the most interesting one.