04 · Industry-Sponsored Project
AnaChart Analyst Accuracy
Making analyst recommendations comparable and measurable
Business value
AUC improved from 0.59 to 0.857, creating a clearer way to compare analyst performance and understand accuracy drivers.
01 · Business problem
55,106 recommendations across 98 NASDAQ-100 stocks used 164 rating labels, making comparison unreliable. The baseline model was barely better than random.
02 · My approach
I standardized the labels, used leakage-aware validation, and engineered stock-difficulty and analyst-momentum features.
03 · What I built
BigQuery analytical tables, systematic label mapping, XGBoost models, analyst segmentation, and a Streamlit prototype.
What changed my thinking
The corrections, tradeoffsand decisions behind it.
- 01
Cleaning created the comparison layer
Reducing 164 labels into five consistent segments was the foundation of the model.
- 02
Accuracy needed context
Analyst performance changes when the stock itself is difficult to predict.
- 03
Recent behavior carried signal
Momentum features added information that long-run averages missed.