Skip to content

Predictor V1 · Model performance

How the model has done.

Live results come first. Older figures are kept apart below and are never added in.

EXPERIMENTAL MODEL

Three separate blocks: the live record, a replay of events before the launch, and the historical validation.

Also on this page: Predictor V3, move size

Live

Live forward performance

Forecasts that were stored before the outcome existed. Since Oct 6, 2026.

Live accuracy

EXPERIMENTAL MODEL

47.7%right: 484 of 1,014 forecasts

95% interval, by trading day: 43.4%–52.3%. With few days it is rough.

ConclusionNo verified edge over the baseline
Total resolved predictions
1,014
Accuracy
47.7%
Majority baseline
52.7%
Edge vs majority
−4.9 pts
Balanced accuracy
49.9%
Up accuracy
52.3%
86 up calls
Down accuracy
47.3%
928 down calls
Trading days
29
Model version
v1-A
Live forecasts and how many were right, by trading day
DayForecastsRightAccuracy
Oct 6, 202610550.0%
Oct 7, 202617635.3%
Oct 8, 202615426.7%
Oct 9, 202610660.0%
Oct 12, 2026371848.6%
Oct 13, 2026402870.0%
Oct 14, 2026533464.2%
Oct 15, 2026592440.7%
Oct 16, 2026512752.9%
Oct 19, 2026481939.6%
Oct 20, 2026693347.8%
Oct 21, 2026411434.1%
Oct 22, 2026261453.8%
Oct 23, 202624937.5%
Oct 26, 2026362363.9%
Oct 27, 2026603050.0%
Oct 28, 2026501836.0%
Oct 29, 2026271140.7%
Oct 30, 2026382155.3%
Nov 2, 2026231356.5%
Nov 3, 2026251768.0%
Nov 4, 2026231147.8%
Nov 5, 2026531834.0%
Nov 6, 2026291241.4%
Nov 9, 2026662740.9%
Nov 10, 2026271244.4%
Nov 11, 202614750.0%
Nov 12, 2026282071.4%
Nov 13, 202615320.0%
  • The majority baseline always calls the more common direction. The stricter of two versions is used: the more common direction among these same events, or among everything known before each day.
  • Edge = accuracy minus the majority baseline, on the same events.
  • Every signal counts once its result is final. Nothing is removed afterwards.

Replay

NOT LIVEEXPERIMENTAL MODEL

Before the live launch

Not live. The same frozen model, run afterwards on what was stored at detection, for events since 2026-10-02 that have no live forecast.

Resolved
129
61 right
Accuracy
47.3%
Majority baseline
50.4%
Edge vs majority
−3.1 pts

129 events before the launch · 0 missed since

Validation

HISTORICALEXPERIMENTAL MODEL

Historical validation

Before 2 Oct 2026. Never added to the live figures.

Accuracy
57.2%
175 of 306 events
Baseline accuracy
52.9%
162 of 306 events
Edge vs baseline
+4.3 pts
Period
Sep 18, 2026 – Oct 1, 2026

Historical. Walk-forward, out of sample, on events detected during the session before 2026-10-02. This is where the model was chosen, so it flatters it.

Scope

Horizons

Only one horizon is an active signal: Session close. The others are stored for research and are never shown as forecasts.

  • 1 hour

    Research only

    Stored for research only: no demonstrated edge

  • 4 hours

    Research only

    Stored for research only: no demonstrated edge

  • Session close

    Experimental signal

    Shown as an experimental signal

  • Next session close

    Research only

    Stored for research only: no demonstrated edge

  • Three sessions

    Research only

    Stored for research only: no demonstrated edge

Rules

How it is judged

When a forecast is written
Written minutes after an event is detected and before its entry time. Never changed afterwards.
What Up and Down mean
Up means the stock does better than the market (SPY) from the entry time to the session close. Down means it does worse.
When a signal is issued
A signal is issued when the event passes the research rules known at detection, was detected during the session, and its prices could be read.
What counts
Every signal counts once its result is final. Nothing is removed afterwards.
The baseline
The majority baseline always calls the more common direction. The stricter of two versions is used: the more common direction among these same events, or among everything known before each day.
The edge
Edge = accuracy minus the majority baseline, on the same events.
The conclusion
Good needs 60% accuracy and an edge of at least 3 points. Below 60% with such an edge is Experimental. Without it: no verified edge. Fewer than 100 live results: no conclusion.

Version

Model

Model
Logistic regression on structured features
Training
Trained on 1,400 events up to Oct 1, 2026
Frozen on
Oct 6, 2026
Version
v1-A
Spec hash
1998b911ac45

The model is frozen: it does not learn from the events it forecasts.

Model performance

EXPERIMENTAL MODEL

Predictor V3 · Move size

A second, separate model. It forecasts the size of a stock's move, not its direction: will the stock move more than it usually does against its own sector fund by the session close? It calls only the clearest cases. It replaced Predictor V2 on Nov 9, 2026. Its numbers are never added to Predictor V1's.

V3 · Live

Live forward performance

Calls that were stored before the outcome existed. Since Nov 9, 2026.

Live accuracy · Predictor V3

EXPERIMENTAL MODEL

66.8%right: 131 of 196 forecasts

95% interval, by trading day: 61.0%–70.8%. With few days it is rough.

ConclusionNot enough live results yet196 of 400 live results needed before a conclusion is drawn.
Total resolved predictions
196
117 moved big
Accuracy
66.8%
Best simple rule
59.7%
Always the more common answer
Edge vs best rule
+7.1 pts
Balanced accuracy
60.5%
Big calls accuracy
65.7%
109 of 166 big calls
Normal calls accuracy
73.3%
22 of 30 normal calls
Trading days
5
Model version
v3-A
The three simple rules, judged on the same live events as Predictor V3
Simple ruleRightAccuracy
Always the more common answer, the best of the three11759.7%
Big if it is already moving a lot today7739.3%
The usual answer for that kind of news11458.2%
Predictor V3 live forecasts and how many were right, by trading day
DayForecastsRightAccuracy
Nov 9, 2026694971.0%
Nov 10, 2026463167.4%
Nov 11, 2026241666.7%
Nov 12, 2026422457.1%
Nov 13, 2026151173.3%
  • Three simple rules are judged on the same events: always the more common answer, big if the stock is already moving a lot today, and the usual answer for that kind of news.
  • Edge = accuracy minus the best of the three simple rules, on the same events.
  • Every call counts once its result is final. Nothing is removed afterwards.

V3 · Validation

HISTORICALEXPERIMENTAL MODEL

Before it went live

The historical test where the model was chosen, then the final check on news nobody had looked at. Never added to the live figures.

Accuracy on its calls
71.4%
364 of 510 calls
Best simple rule
57.6%
The usual answer for that kind of news
Edge vs best rule
+13.7 pts
Period
Sep 10, 2026 – Oct 21, 2026

Historical. Walk-forward, out of sample. This is where the model was chosen, so it flatters it.

Final check: accuracy
78.2%
169 of 216 calls
Edge vs best rule
+3.7 pts
Result
Passed
Period
Oct 22, 2026 – Nov 4, 2026

V3 · Rules

How it is judged

When a forecast is written
Written minutes after an event is detected and before its entry time. Never changed afterwards.
What Big and Normal mean
Big means the stock moves more than its usual move against its sector fund, from the entry time to the session close. Normal means less.
When a call is made
Predictor V3 makes a call only when its inputs point clearly one way (about 4 in 10 events in its historical test). The rest are marked No call.
What counts
Every call counts once its result is final. Nothing is removed afterwards.
The three simple rules
Three simple rules are judged on the same events: always the more common answer, big if the stock is already moving a lot today, and the usual answer for that kind of news.
The edge
Edge = accuracy minus the best of the three simple rules, on the same events.
The conclusion
Good needs 60% accuracy and an edge of at least 3 points. Below 60% with such an edge is Experimental. Without it: no verified edge. Fewer than 400 live results: no conclusion.
Predictor V2
Predictor V3 replaced Predictor V2 on the site from its live start. Predictor V2 keeps running unseen until V3 has 400 results, so the two can be compared on the same news; then it stops.

Experimental research model. A forecast is a statistical statement about past patterns, often wrong, and not investment advice. ShockSense shows what the model said and what happened. It does not tell anyone what to do.