HN Showcase
Back to browse
I logged Gemini's stock predictions for 38 days to study LLM drift

I logged Gemini's stock predictions for 38 days to study LLM drift

by clsia·Mar 8, 2026·5 points·1 comment
Visit ProjectView on HN

AI Analysis

●●SolidBig BrainRabbit HoleNiche Gem

Rigorous 38-day Gemini drift study with citation-mapped predictions and confidence scores.

Strengths
  • •Mandatory-citation tracking in prompts reveals where LLM sourcing comes from vs. hallucination
  • •Look-ahead-bias-free methodology (predictions use only prior data) is structurally sound
  • •Calibration rubric (0-1 confidence scale) enables quantitative drift analysis across checkpoints
Weaknesses
  • •Single model (Gemini 2.5) limits generalization—doesn't compare drift across Claude, GPT, etc.
  • •Dataset is small (1.1K rows) and focused only on 38 days of stock predictions
Category
Data
Target Audience

ML researchers studying LLM reliability, financial bias, and temporal drift in foundational models

Similar Projects

AI/ML●●Solid

Axon – Agentic AI with mandatory user approval and audit logging

Agent approval gates and audit logs beat open-source alternatives, but multi-agent governance isn't novel.

Solve My ProblemBig Brain
NeuroVexon
123mo ago