Skip to content

About

A research tool for how markets react to news.

ShockSense studies a narrow question carefully: after a financial headline reaches the system, what did the market do next — and how often did similar events move the same way?

The problem

“Shares jumped on the news” hides three questions.

When was the news actually known? Articles are often published, updated and syndicated hours apart. A move that happened before anyone could read the headline is not a reaction to it.

Was it the news, or the market? On a day the whole market rises, most stocks rise with it. A reaction needs a benchmark.

Does it happen reliably? One or two examples prove nothing. Without enough comparable events, the honest answer is that we don't know yet.

What it does

Four steps, in order.

  1. 01

    Classifies

    Turns a headline into a structured event: category, company, ticker and a classifier confidence.

  2. 02

    Measures

    Records the market movement after ShockSense detected the headline, raw and adjusted for SPY.

  3. 03

    Compares

    Places a new event next to every eligible prior event of the same kind, observation by observation.

  4. 04

    Abstains

    Reports a historical bias only when fixed evidence thresholds are met — and says so when they are not.

What it is not

  • A trading signal or a recommendation to buy or sell anything.
  • A forecast, with one labelled exception. The historical statistics and the lean describe the past; they are not forecasts. The one forecast on the site is the Predictor, which is experimental and labelled as such.
  • Investment advice of any kind.

Architecture

Two paths that never mix.

Research events enter only through the trusted ingestion pipeline. The public site and analyzer read that data but never write to it, so what visitors type can never change the evidence.

Research path · trusted

News source

Marketaux financial news, with publication time

Ingestion pipeline

Already seen → staleness → noise screen → AI classification → ticker validation → duplicates. Every rejection recorded.

Measurement

Post-Detection Return and SPY-adjusted return, once each window has closed

Research database

PostgreSQL · events, reactions, eligibility flags, run logs

Public path · read-only

ShockSense web

This site. Receives only browser-safe configuration.

Public API

Rate-limited, bot-checked. Reads eligible events and patterns; returns an allow-listed set of fields.

Analyzer

AI classification (server-side only), then a comparison with the eligible history.

Analysis cache

Identical headlines are answered once. Never written into the research data.
API keys and the database connection live only on the server. The browser talks to the public API, never to the database or to any AI provider.

Pilot status

Validated method, small sample.

Raw observations
4105
Eligible events
3128
Historical biases issued
6

The measurement method, eligibility rules and evidence thresholds are complete and tested. The pilot dataset is deliberately small, so ShockSense currently abstains on every category. New events are added through the hardened pipeline; the evidence grows only as fast as real events accumulate.

Built with

The building blocks.

Classification
ShockSense classification model, through a fixed schema
News
Marketaux financial news
Market data
Yahoo Finance prices; NYSE calendar via exchange_calendars
Backend
Python, FastAPI, SQLAlchemy, PostgreSQL
Frontend
Next.js, React, Tailwind CSS, Motion
Typography
Geist and Instrument Serif

Historical market-reaction analytics for research and educational purposes. Not investment advice.