Shipped · Research pipeline

FinSight

An earnings-intelligence research pipeline — from 14,584 raw transcripts to a walk-forward alpha signal.

ProblemEarnings-call language is full of sentiment signal, but most of it doesn't survive honest, look-ahead-free testing.

ApproachParse 14,584 transcripts into sentiment and retrieval-based features, then score every signal only on data it could have seen at trade time.

ResultA real signal survives in the energy sector — everywhere else, the edge is reported as thin rather than inflated.

SignificanceHonest reporting of where a signal does and doesn't work is rarer in quant research than the signal itself.

14,584Earnings transcripts

+0.31Energy-sector IC

FinBERT+ RAG features

601Companies covered

The System

14,584 earnings-call transcripts are parsed into FinBERT sentiment and RAG-retrieved contextual features, then fed into a strict walk-forward backtest so every signal is scored only on data it could have seen at trade time.

The Result

The pipeline surfaces a genuine cross-sectional signal in the energy sector — an information coefficient of +0.31 — while honestly reporting where the edge is thin. The findings ship as a live Next.js 14 analytics dashboard, not a static notebook.

Why It Matters

FinSight is the deployment counterpart to my gated-trading research: the same walk-forward discipline that refuses false alpha, packaged as a tool a user can actually open and explore.

Methodology Pipeline

Key Result

Energy sector+0.31 IC

Other sectorsweaker / not significant

The edge is real, but it doesn't live everywhere — and the pipeline says so

↗ View live dashboard → See the explorable version