Live research system · 29 Jun 2026 → 10 Aug 2026

We measure what news
actually does to prices.

QuantCannon reads financial news as it breaks, scores it, and then tracks what the market did next — down to the tick. The result is a measurement layer: which sources are predictive, which stocks a story really touches, and how often the model is wrong.

We publish the misses alongside the hits. Right now the strategy is not beating a random-entry baseline, and you can read that number on this page.

See the evidenceHow it works
Signal tape
measured · 1d outcome
  • 10:18Finnhub

    3 Big Takeaways From Microsoft's Earnings

    MSFTscore 8▲ up+4.87%HIT
  • 13:42Yahoo Finance

    Marriott Outlook Comes Up Short As Middle East Conflicts Drag On Results

    MARscore 7▼ down-2.98%HIT
  • 14:10Investing.com

    Amazon surpasses $3 trillion valuation as AI and cloud propel shares

    AMZNscore 8▲ up-2.1%MISS
  • 12:23PR Newswire

    Realty Income Receives 'A' Credit Rating from Fitch Ratings

    Oscore 7▲ up-2.26%MISS
  • 13:54Finnhub

    Tesla: Record Deliveries Do Not Fix The Return On Capital Problem

    TSLAscore 8▼ down+1.74%MISS
  • 14:05Finnhub

    Nvidia's Pause Before The Next AI Wave

    NVDAscore 7▲ up+3.96%HIT
Real signals from the research database. Direction is what the model predicted; the outcome is what the market did one day later.
10,770
Measured signals
each with 1h → 5d outcome windows
657,790
Price snapshots
15-minute path reconstruction
2,773
Tickers observed
from a 5,268-name universe
11,756
Logged decisions
every trade and every rejection, with reason
Source reliability

Not every headline is worth the same

Directional accuracy per news source, measured on signals where a direction was predicted and the outcome window has closed. Sources with fewer than 30 measured signals are excluded.

  • Finvizn=109
    61.5%
  • MarketWatchn=60
    60.0%
  • PR Newswiren=282
    56.7%
  • Investing.comn=159
    54.1%
  • GlobeNewswiren=574
    53.0%
  • Finnhubn=1,213
    51.0%
  • Yahoo Financen=685
    49.2%
Most predictiveAt or below a coin flip

A coin flip is 50%. Two of the six sources we ingest sit at or below it — which is exactly the kind of finding that stays hidden when a platform only advertises its winners.

Evidence, unedited

Here is where the system currently fails

Paper-trading results since the execution layer was rebuilt on 30 July 2026. The sample is small and we say so: twenty-four closed trades prove nothing yet. We publish it anyway, because a platform that only shows numbers once they look good is not a measurement platform.

Signal quality vs. random entry

Over the same window, in the same tradeable universe, entering a random news-carrying stock beat our own selection over five days.

  • Random entry, same universe+3.82%
  • QuantCannon selection-1.18%
−5%0+5%

We traced the likely cause: entries were landing an average of 0.55% above the price at signal time — buying the spike rather than the move. Entry thresholds were tightened on 8 August and the configuration is now frozen until 100+ trades accumulate.

24
Closed trades
too few to conclude anything
37.5%
Win rate
profit factor 0.91 — below 1.0 means losing
−$31.29
Net result
on a $10,000 simulated book
Which exit rule actually made the money

Every closed position, grouped by the rule that closed it. The uncomfortable part is the top row: more of the gains come from positions simply reaching their holding limit than from the strategy's own profit-taking. On this evidence the system is currently better at being patient than at being selective.

  • Holding period expired11 trades+$79.05
  • Stop-loss7 trades−$284.42
  • Take-profit3 trades+$109.98
  • Trailing stop2 trades+$8.70
  • AI review1 trade+$55.40
net profit & loss per exit rule
Model confidence vs. reality

If the AI score carried clean information, accuracy would climb with it. It does not — the 8 bucket outperforms the 9–10 bucket, and the top bucket is thin. This is an open problem, shown rather than smoothed. Dashed line marks a coin flip.

Score 9–10n=68
51.5%
Score 8n=749
57.7%
Score 7n=1,594
49.7%
Below 7n=1,737
50.5%
What we built

A measurement layer, not a tip service

Three things come out of the same pipeline. None of them require you to trust a track record.

01 · Impact engine

Which stocks does this story move?

A headline goes in; out comes the directly named ticker plus the second-order names it tends to drag with it — suppliers, competitors, sector peers — with an expected direction and a horizon.

02 · Source scoring

How much is this outlet worth?

Every source carries a measured, continuously updated hit rate per catalyst type. Trust is earned from outcomes, not assigned by reputation.

03 · Open audit trail

Every decision, with its reason

Trades, rejections, stop-outs and the rule that fired are all logged and queryable. Rejections are the interesting half: 11,756 decisions recorded so far.

How a signal is produced

  1. 1
    Ingest

    Four news feeds polled on a session-aware schedule.

  2. 2
    Filter

    Non-financial noise and duplicates dropped before any model runs.

  3. 3
    Score

    Two independent passes: a language model on the text, pure statistics on the market data.

  4. 4
    Gate

    A trade opens only when both passes agree, or one is exceptionally strong and the other does not veto.

  5. 5
    Measure

    Outcome tracked at 1h, 4h, 1d, 3d and 5d, then written back as source trust.

Early access to the impact API

We are looking for a small number of design partners — fintech products, brokerages and financial media — who want structured news-impact data rather than another signal feed. Tell us what you would query and we will tell you honestly whether our data can answer it yet.

Get in touchquandcannon@gmail.com