Skip to content

Methodology

How ShockSense measures a reaction.

A reaction is only meaningful if it is measured from a moment ShockSense could actually have acted on, against a clear benchmark, on a sample large enough to say anything. This page describes each rule, and where the current pilot data stops.

01

Principles

ShockSense is a research tool, not a signal service. Four rules shape every number on this site.

Detection is the anchor

Returns are measured only from after ShockSense received the headline, never from the earlier publication time.

Every observation is traceable

Excluded events are flagged with a reason, not deleted. The funnel from raw data to the analysis set is published.

Abstain by default

A historical bias is reported only when fixed thresholds are met. Otherwise the answer is “Not assessed” or “No clear signal”.

One calculation layer

The research console, the API and this site all compute statistics with the same shared module.

02

From headline to event

New events enter through a single trusted path. Each stage can reject a headline, and every rejection is stored with its stage and reason. Headlines typed into the public analyzer never take this path and are never added to the research data.

  1. Headlines arrive

    Financial news from Marketaux, checked every hour of the trading day. ShockSense stores the provider's publication time and, separately, the moment it received the headline — its detection time.

  2. 01Already seen

    Articles stored or finally rejected before — matched by URL and by exact or normalized headline — are skipped, so nothing is classified twice.

  3. 02Publication time

    A missing publication time can never make an article stale; a time slightly after detection is treated as clock drift.

    Rejects: Timestamps that cannot be read as a point in time

  4. 03Staleness

    Publication and detection times are compared on the exchange calendar, before any AI call.

    Rejects: A full trading session passed between publication and detection

  5. 04Noise screen

    Word-bounded rules, also before any AI call.

    Rejects: Recaps, transcripts, market wraps, price-move reports, previews and opinion

  6. 05Classification

    Our classification model returns the category, primary company and ticker through a fixed schema.

    Rejects: Not a market event · several companies with no clear subject · no listed ticker

  7. 06Confidence

    The classifier must be reasonably sure of the category.

    Rejects: Classifier confidence below 72%

  8. 07Ticker validation

    The ticker must exist today and belong to the named company. An outdated symbol is replaced only when a name search finds exactly one matching US listing; market proxies follow a fixed policy.

    Rejects: A ticker that belongs to a different company · non-equity listings · proxies outside the policy

  9. 08Duplicates

    Compares against recent events with the same subject and category.

    Rejects: A near-identical headline about the same subject within five days

  10. Stored as a live event

    Its reaction is measured only after the measurement window has closed. Every rejection is kept with its stage and reason, so the funnel can always be audited.

The ingestion pipeline: each headline passes eight stages in order; any stage can reject it with a recorded reason.

03

Classification

Our classification model reads each headline and returns a structured record through a fixed schema: the event category, the primary company, its ticker and a confidence score. Headlines are passed to the model as quoted data, not instructions. When a headline names several companies without a clear subject, or no listed company at all, the classifier abstains instead of guessing.

Market-wide events are measured on a small, fixed set of ETF proxies — SPY for the US equity market, USO for crude oil, GLD for gold and EWZ for Brazilian equities — each allowed only for specific categories. Classifier confidence describes how sure the model is of the category. It is not a probability that the market will move.

News subtypes. The four company categories also get a subtype that says which way the news cuts for that company: earnings (guidance raised, guidance cut, results beat, results miss), M&A (company being acquired, company buying, deal blocked or ended), regulation (approval or legal win; fine, probe or legal loss) and company news (share or debt offering, buyback or dividend raise, dividend cut, contract or partnership, operational setback, layoffs or restructuring, leadership change, bankruptcy). Anything else is “other”. The model assigns the subtype from the headline text alone — never from a price — and stored events were labelled the same way. Subtypes are used by the model portfolio; the pattern pages stay per category.

Earnings
A company's reported financial results, or a change to the company's own financial guidance/outlook (raise, cut, withdrawal, pre-announcement).
M&A
Mergers, acquisitions, divestitures, takeover bids, and definitive deal agreements, approvals or completions involving a specific company.
Monetary Policy
Central bank rate decisions and official monetary-policy statements (Federal Reserve, ECB, Banco Central do Brasil, …).
Fiscal Policy
Government budget, taxation, spending, tariff and trade-policy actions.
Inflation / Macro
Official macroeconomic data releases (CPI, PCE, payrolls, GDP, retail sales, PMIs).
Geopolitics
Wars, sanctions, elections and diplomatic developments with market-wide implications.
Commodity Shock
Supply or demand shocks to commodities (producer decisions such as OPEC quotas, supply disruptions).
Regulation
Regulatory, legal or enforcement actions against or affecting a specific company or sector (approvals, fines, lawsuits, antitrust rulings).
Company-specific
Other discrete company events: management changes, product launches or recalls, major contracts, capital actions (buybacks, offerings, dividend changes), bankruptcy filings.

04

Published vs. detected

Every event carries two timestamps. Published is the provider's time for the article, kept for provenance. Detected is when ShockSense received it — the only moment from which a reaction can honestly be measured, because anything earlier was not observable to the system.

If the market has already traded a full session on the news before ShockSense sees it, the move after detection measures a re-report, not a reaction. Those articles are rejected as stale.

P Published (provider time)D Detected by ShockSense Regular session, 09:30–16:00 ET
  • P Fri 18:00D Mon 08:00

    Kept. No full session traded between the two times.

  • P Mon 10:00D Tue 15:00

    Kept. Monday's session was already open at publication.

  • P Mon 08:00D Tue 08:00

    Rejected as stale. The market traded all of Monday on the news first.

An article is stale when a complete regular session lies entirely between publication and detection. Holidays and early closes come from the exchange calendar. When the provider gives no publication time, the rule cannot be applied and the timestamp is stored as unknown — never replaced.

Backfilled history. To give every category a starting sample, the month before the model portfolio's launch (August 25 to September 23, 2026) was read from the news archive after the fact and run through the same pipeline and rules. Those articles were never seen live, so their detection time is set to their publication time plus one hour — about when the hourly job would have picked them up. They are labelled Backfill wherever they appear, count as evidence, and are never traded by the model portfolio.

05

Post-Detection Return

The Post-Detection Return runs from the first eligible market observation at or after detection to the official close of the first full trading session after detection. All session logic uses the New York Stock Exchange calendar, including holidays, early closes and daylight saving.

  1. Published

    Provider timestamp. Stored for provenance, never used as the anchor.

  2. Detected at

    When ShockSense received the headline. The only anchor.

  3. First eligible observation

    Open of the first hourly bar at or after detection — or the next official open.

  4. Closing observation

    Official close of the first full session after detection.

  • During the regular session

    Start: Open of the first hourly bar starting at or after detection (searching forward up to two hours, never backward).

    End: Official close of the next session.

  • Before the open, or exactly at it

    Start: That session's official open.

    End: Official close of the same session.

  • Late in the session, after the close, on a weekend or holiday

    Start: The next session's official open.

    End: Official close of that session.

Returns use actual traded prices, with dividends added back when they go ex inside the window. A stock split inside the window, a missing price or a mismatch between hourly and daily data makes the measurement unavailable — missing data is never turned into 0%. A window that has not closed yet is reported as pending.

06

Market adjustment

The raw return shows what the asset did. To separate the event from what the whole market did at the same time, ShockSense also reports a market-adjusted return: the asset's return minus SPY's return over the identical start and end observations.

Definition

adjusted = asset − SPY

Same timestamps, same price types. It assumes the asset moves one-for-one with the market — a deliberately simple control, not a factor model. It is undefined when the measured asset is SPY itself.

Example · MAC · M&A

MAC raw
−1.93%
SPY, same window
−1.29%
Market-adjusted
−0.64%

07

Eligibility

Not every stored observation belongs in the analysis. Each event is checked against the rules below, in order, and keeps every reason that applies. Nothing is deleted: excluded events stay in the database with their flags, and the final manual review of the pilot set is kept under version control with a reason for each decision.

StepRemovedRemaining
  1. Raw observations in the pilot database

    4105

  2. Hand-written sample events, not live detections

    −12

    4093

  3. Recaps, transcripts, market wraps and price-move reports

    −187

    3906

  4. Failed the event-quality rules

    −55

    3851

  5. Ticker did not belong to the company in the headline

    −3

    3848

  6. Excluded in the manual pilot review (stale, ambiguous or not an event)

    −11

    3837

  7. No valid Post-Detection measurement

    −204

    3633

  8. Duplicate of an earlier event

    −505

    3128

  9. Validated pilot events

    3128

08

Evidence and abstention

A category shows a historical bias only when all four gates pass on its eligible history. The gates are fixed in one place and apply identically to raw and market-adjusted returns.

Sample size
n ≥ 5
Consistency
≥ 70%
share moving in the majority direction
Magnitude
|mean| ≥ 0.5%
Agreement
mean, median
same sign as the majority

Generic earnings is always reported as “No clear signal”, even with enough history: the category mixes beats and misses, so a single direction for “earnings” would describe the sample, not the news. Its statistics are still shown.

Historical bias

  • Positive historical bias

    At least 5 prior events, at least 70% moved up, the mean move is at least +0.5% and the median agrees.

  • Negative historical bias

    The same gates, in the downward direction.

  • No clear signal

    Enough history to assess, but the reactions were mixed, small, or the category is too broad to carry a direction (generic Earnings is always reported this way).

  • Not assessed

    Fewer than 5 eligible prior events. ShockSense abstains rather than guess.

Evidence level

  • Insufficient historyFewer than 5 eligible events.
  • Early evidence5–19 eligible events.
  • Established sample20 or more eligible events.

09

Walk-forward check

To check the rules without hindsight, every eligible event is judged as if in real time: its history is only the same-category events whose measurement windows had closed before it was detected. A signal counts as a hit only if the event then moved in the stated direction.

Bars run from detection to the close of each event's measurement window. Only windows that closed before the judged event was detected are used, so no result ever depends on information from its own future. Schematic, not data.

Raw returns

6 / 3128

events where a historical bias was issued

Market-adjusted

7 / 3034

events where a historical bias was issued

On the pilot sample no category reaches the thresholds at any point, so no bias is issued and there is no hit rate to report. That is the expected result for a dataset this small, and it is reported as it is.

10

Model portfolio methodology

The ShockSense Model Portfolio asks one question: if ShockSense signals had been followed prospectively, how would a simulated portfolio have performed? The current version (methodology 1.2) started with $100,000 of simulated cash at the market open on September 28, 2026, follows fixed rules set before its first decision and never places an order. No real money is invested. It is not backfilled: earlier events, including the labelled one-month backfill, count as evidence for its signals, but no position exists before its start.

  • Signals

    Evidence per news subtype (for example Earnings · Guidance raised), on market-adjusted returns, taken at each event's detection time and frozen there. A subtype qualifies when at least 30 earlier events had closed windows, one direction beat a coin flip in an exact one-sided sign test (p ≤ 0.05), the average move was at least 0.5%, and mean and median agreed. “Other” subtypes and the market-wide categories never trade. A later event, label or price can never change a decision already made.

  • Which events

    Live events detected from the start onward that pass every research rule knowable at detection. A decision waits until the event and every earlier measured event of its category carry a subtype label. Headlines typed into the analyzer never become events, so they can never create a position.

  • Entry and exit

    Entry is the event's first eligible post-detection price, exactly as the measurement defines it. Exit is the measurement end: the official close of the exit session. No stops, targets or discretion.

  • Size and limits

    Each position is 10% of equity at entry (open positions counted at cost). Gross exposure never exceeds 30% of equity, and a ticker is held at most once. No stock priced below $5 is entered: such prices move too violently for a position sized like the others. A signal that breaks a limit, or has no valid entry price, is recorded as not allocated with its reason.

  • Costs

    Initial simulation excludes transaction costs and slippage (0 bps). Both are stored as parameters, so a later version can include them without rewriting history.

  • Benchmark

    SPY bought with the same $100,000 at the official open of the inception session, with dividends reinvested at the close of their ex-dividend session, on actual traded prices.

Returns

asset return
r = (exit + dividends) / entry − 1
simulated long
r
simulated short
−r = 1 − (exit + dividends) / entry
value
allocation × (1 + position return)

A simulated short of a fixed dollar amount gains what the asset loses, so its return is the asset return with the sign flipped. The ratio entry / exit − 1 is not used: it overstates gains and understates losses. The short's allocation is held aside as collateral while it is open; its proceeds are never counted as spendable cash.

Why the rule changed. Version 1.0 (September 24–27, 2026) used the site's category-level 70% consistency rule. It made no trades: once news is out, most of the move happens before a system that reads it can enter, and what remains is a smaller, noisier follow-through that no broad category repeats 70% of the time. The 70% rule is right for the public labels, which should not call a 56/44 split a pattern; it is the wrong test for a strategy. Version 1.2 tests sharper news types statistically instead. Several types are tested, so a qualifying type can still be luck: the forward record, judged only on events after September 28, is the test. The 1.0 record is kept unchanged. The $5 price floor was added on September 29; the two positions opened below $5 on September 28 were removed as if never placed.

The equity curve has one point per completed trading session, written once every position that could affect it is known and never rewritten, plus a latest point refreshed each hourly run. A position whose exit cannot be measured stays open and is held at cost rather than given an invented exit. Portfolio methodology version 1.2.

11

The lean

Most events never earn a historical bias, and the model portfolio acts on few of them. The lean answers the smaller question that is left: which way did earlier events of the same kind go? It is reported for every event, however weak, and says how weak.

  • Similar events

    Events of the same category and, for company news, the same news subtype (for example Earnings · Guidance raised; “Other” is its own group). Market-wide categories have no subtypes, so the category is the group.

  • Direction

    The majority direction of those earlier events, measured against the market (the market-adjusted Post-Detection Return). With fewer than 10 earlier events, or an even split, there is no lean.

  • Strength

    How unlikely a fair coin is to be that one-sided, in an exact one-sided sign test: strong at p ≤ 0.05, moderate at p ≤ 0.20, weak otherwise. A weak lean is close to a coin flip and is labelled that way.

  • Events measured on SPY

    An event measured on the benchmark itself has no market-adjusted return. Its lean uses the raw move of earlier events that were also measured on SPY.

  • Stored at detection

    For a live event, only events whose window had closed before its detection time count. The hourly job stores the lean once; later events, labels and prices never change it.

  • Scoring

    After the close that ends the event's window, the lean is right if the event moved that way against the market. The scorecard counts live events detected from the market open on October 2, 2026, when this rule was fixed. Weak leans are scored too, and should sit near 50%.

The lean changes neither the historical-bias labels above (the 70% rule) nor the model portfolio's rule; all three read the same measurements with different bars. It is not a forecast: a lean describes what earlier events did, and the scorecard shows how often that carried over. In the analyzer, a headline's lean uses everything measured so far and is never scored. Lean rule version 1.0.

12

Predictor V1 (experimental)

Predictor V1 is the first of the two forecasts on this site. For each company news event detected during the trading session, it says whether the stock will do better or worse than the market (SPY) from the entry time to the session close. It is experimental, and it is labelled that way wherever it appears.

  • What it forecasts

    Direction only. Up means the stock does better than SPY from the entry time to the session close. Down means it does worse. There is no size and no price target.

  • A frozen model

    One model, trained on events detected before 2 Oct 2026 and then frozen. It does not learn from the events it forecasts.

  • Written before the outcome

    A forecast is written minutes after the event is detected and before its entry time. It is never changed afterwards.

  • When a signal is issued

    The event must pass the research rules known at detection, be detected during the session, and have prices that can be read. Otherwise the event gets no signal, and the reason is shown.

  • No score is shown

    The model's score is not shown, because it is not a calibrated confidence. Only the direction is published.

  • The live record

    Every signal counts once its result is final. Nothing is removed afterwards.

  • The baseline

    The majority baseline always calls the more common direction. It is the bar the model has to beat.

  • The edge

    Accuracy minus the baseline, on the same events, in percentage points.

The conclusion follows a fixed rule. Good: at least 60% accuracy and an edge of at least 3 points. Experimental: below 60% with such an edge. Otherwise: no verified edge. With fewer than 100 live results, no conclusion is drawn.

Live results, the replay of events before the launch and the historical validation are shown in separate blocks and are never added together. A forecast is often wrong, and it is not investment advice.

Predictor V2 (experimental) is a second, separate model. It forecasts the size of the move, not its direction: Big when the stock moves more than it usually does against its own sector fund by the session close, Normal when less. Its live forecasts cover events detected from Monday, October 26, 2026. It is judged against three simple rules on the same events, and no conclusion is drawn before 400 live results. Its record is kept apart from V1's.

13

Limitations

  • Small sample

    The frozen pilot holds nine eligible events detected over five days; the one-month backfill and the live pipeline add to it, but most categories are still far from established patterns. Backfilled detection times are an approximation (publication + 1 hour).

  • Hourly entry grid

    Entry can be up to about an hour after detection (two when a bar is missing). The delay is recorded for every event.

  • Regular session only

    Pre-market and after-hours trading are not used for measurement.

  • US listings

    Measurement currently covers US-listed securities and a few US-listed ETF proxies.

  • Simple market control

    The adjusted return assumes a beta of one against SPY; it is not a risk model.

  • Classifier errors

    The model can misread a headline. Validation catches wrong tickers, not every wrong category.

  • One news source

    New headlines come from Marketaux, which does not cover every outlet. Each hourly run takes the newest articles, so busy hours are sampled. Late articles are rejected as stale, never measured.

  • Association, not prediction

    Past reactions describe history. They are not a forecast and not investment advice.

14

Research history

ShockSense started as a prototype that classified headlines and presented directional calls with expected one-day moves. An audit of that prototype changed the project.

  1. Prototype

    Headlines were classified by an AI model and scored against a publication-time anchor. Missing publication times were silently replaced with the current time, and hand-written sample events sat alongside live ones.

  2. Audit

    The review found timing that could credit moves from before the system saw the news, recaps and transcripts counted as events, tickers that belonged to the wrong company, and directional calls drawn from tiny samples.

  3. Rebuild

    Publication and detection were separated, every event was re-measured from detection with a market-adjusted variant, exclusions became auditable flags, and one statistics module replaced ad-hoc calculations. Directional call language was removed.

  4. Frozen pilot

    A documented manual review left nine eligible events. The walk-forward check issued no historical bias — and the product was designed to say so plainly.

  5. Now

    A hardened ingestion pipeline adds new events with both timestamps, validated tickers and staleness checks, on top of a labelled one-month backfill run through the same rules. The evidence grows only as fast as real events accumulate.

15

Versions

Every published number is tied to the versions that produced it.

Methodology
1.2
Measurement
post-detection-v1
Analytics
v1
Extractor
v2
Benchmark
SPY