---
title: "AI agents boost econometric coding accuracy to 97%"
slug: "ai-agents-boost-econometric-coding-accuracy-to-97"
published: "2026-10-20"
beat: "Research"
tags: ["Research"]
creator: "Agentry Newsroom"
editor: "Susanne Sperling, Editor — Human in the Loop"
tools: ["Claude (Anthropic)", "Perplexity Sonar"]
creativeWorkStatus: "verified"
dateReviewed: "2026-10-20"
aiActArticle50: "compliant"
humanView: "https://agentry.news/research/ai-agents-boost-econometric-coding-accuracy-to-97"
agentView: "https://agentry.news/agent/ai-agents-boost-econometric-coding-accuracy-to-97"
---# AI agents boost econometric coding accuracy to 97%

> A benchmark published October 18–19, 2026 tested 40 econometric tasks across Stata, R, and Python, finding that agent workflows executing and self-repairing code increased correct results from 79% to 

*Drafted by an AI agent. Verified by Susanne Sperling, Editor — Human in the Loop. [AI policy](/ai-policy).*

A benchmark of econometric coding tasks published by the Centre for Economic Policy Research found that AI agents substantially improve accuracy when permitted to execute and revise code, according to a [VoxEU column](https://cepr.org/voxeu/columns/letting-ai-agents-write-econometric-code) dated October 18, 2026, with an update on October 19.

The study evaluated **40 econometric tasks** covering linear regression, difference-in-differences, instrumental variables, regression discontinuity, discrete choice, survival analysis, and time series across Stata, R, and Python. The benchmark compared models generating code without execution against workflows that ran code and enabled self-correction. Results counted as correct when reported numbers matched precomputed reference outputs within a specified tolerance, based on simulated datasets.

## Agent Workflows Outperform Chatbot Runs

Across 14,400 total runs, the benchmark found that **execution and self-repair increased the overall correct-result rate from 79% to 97%**. Model-specific results showed substantial gains: [Claude Sonnet 4.6 in Claude Code achieved 81.4% accuracy in chatbot runs, 93.5% with self-repair, and 99.2% in agent workflows](https://cepr.org/voxeu/columns/letting-ai-agents-write-econometric-code). GPT-5.6 Luna in Codex achieved 77.3%, 90.3%, and 95.1%, respectively.

The finding underscores a key distinction in agent design: **agentic workflows that iterate on code execution systematically outperform single-pass generation**. This aligns with broader industry patterns in which agent frameworks enable models to test, observe failures, and adapt in real time rather than committing to a single output.

## Implications for Economic Research

Econometric analysis—regression, causal inference, time series forecasting—depends on precise code execution and careful interpretation of statistical output. The benchmark demonstrates that current large language models paired with execution environments can handle these tasks at rates competitive with hand-written or expert-reviewed code, provided the agent architecture includes a feedback loop.

The study used simulated datasets and reference outputs, meaning results reflect controlled conditions rather than real-world data discovery or model specification challenges. Nevertheless, the scale of improvement (79% to 97%) suggests agent-assisted econometric workflows may reduce errors in standard statistical tasks, a claim relevant to academic research, policy analysis, and financial modeling.

## Next Steps

The results suggest that agentic coding frameworks deserve scrutiny in domains where correctness is high-stakes—finance, epidemiology, regulatory analysis. Future work may examine how agents perform on novel datasets, non-standard econometric tasks, or code review and interpretation, areas where execution alone does not guarantee validity.