Loading…
Loading…
Start here — the mental model
One question: should a human investigate this market now?
What the score means
Notice the alarm
Score ≥40, strict Level‑2, or several sibling markets moving together means “stop and inspect”.
Look for public explanation
Check news, official statements, and whether the market is moving only because the deadline is near.
Group related markets
Ten Iran rows are one Iran alarm family, not ten independent proofs.
Escalate only if unexplained
If high score + no obvious public reason + multiple related markets: send a heads-up for human review.
Shortest version: “Market probability jumped + real money/trades arrived + buyers were mostly on Yes” is interesting, but only after the red-team checks: false positives at looser thresholds, correlated sibling markets, near-certainty settlement effects, and public-news explanations.
Start with current markets
Active alerts right now
Loading…
What the current analysis says
Summary metrics
Proposed signal levels
- Level 1 / monitor: unusual volume plus price lift. Good for “something is moving; check news/source quality now.”
- Level 2 / informed-flow: Level 1 plus positive Yes-side pressure and not-yet-certain price. In the current sample this fired rarely but cleanly.
- Context filter: group sibling markets, down-rank scheduled deadline/settlement arbitrage, and snapshot public news at alert time.
Honest interpretation of “9/10”
The backtest produced — for the strict rule and — for a softer rule. That is stronger than 9/10 in this dataset, but it is not yet an out-of-sample reliability claim because many rows are sibling markets around the same events.
Reality check / red-team view
Where the signal is fragile or misleading
Looser thresholds create many false positives
Stacked bars: green = true event, red = false alarm. This is why “watchlist” is not “prediction”.
Default alerts are clustered, not independent
Rows selected by the current score threshold, grouped by coarse topic family.
Plain-language caveat
Loading…
How to avoid fooling ourselves
- Look at false positives just below the default threshold, not only the clean high-score rows.
- Count topic families before counting market rows; Iran sibling markets are highly correlated.
- Treat near-certainty/deadline markets as possible settlement mechanics, not early warning.
- Require an external-news snapshot when a live alert is escalated.
Examples the simple story hides
Metric explainer
What exactly is the alarm score measuring?
Score formula used in the dashboard
score = price level + price lift + relative lift + flow surge + flow size + net Yes pressure
The market percentage itself is p(t): Yes price ≈ market-implied probability. The “derivative” part is Δp over a fixed window, mostly p(−6h) − p(−24h). Because that window is 18 hours, its slope proxy is Δp / 18h. The score also uses cash-flow features, because price-only jumps can be stale or deadline mechanics.
Score components on real rows
Stacked bars show which inputs created the 0–100 score. Red rows are false alarms / noisy rows.
Used vs not used
Important interpretation
Score 40 does not mean 40% probability. It means enough unusual things happened together to justify human review. A market can be 60% because everyone already knows the news; that is different from a 10%→35% move with new Yes-flow and volume surge.
Probability over time
Market percent, derivative proxy, and early-warning horizon
Probability paths before resolution
Faint lines = all resolved rows with data; thick lines = illustrative high-score / false-positive examples.
Derivative proxy: percentage-points per hour
Each dot is Δp from −24h to −6h divided by 18 hours. Larger bubbles have more 6h flow.
If I want warning hours/days before, how noisy is raw probability?
Drag the lead-time slider. Raw probability here means only the market Yes price at that snapshot; it does not use volume, account history, time-decay, or a model formula. Bars and KPIs show TP/FP/FN trade-offs for that simple price threshold.
Answer for that warning lead
Loading…
If I only need ~2 hours of heads-up: false positives and misses
Loading…
Concrete error examples: negative false positives and Yes misses
Same event, different deadlines
Term structure: by-date markets are not interchangeable
Why this matters: “Event by May 24” and “event by June 15” can both be about the same story but carry different cumulative probabilities. Convert them to comparable per-day hazard to see whether the market thinks risk is front-loaded or just spread over more time.
Cumulative event probability by deadline
For survival-style markets, this chart uses event risk = 1 − Yes price.
Per-day hazard and theta-adjusted move
Hazard normalizes different deadlines; theta-adjusted move asks whether probability rose beyond mechanical time decay.
Rows behind the selected term curve
Alarm loudness
Choose how much activity is enough to alert
If I set the alarm here, what happened historically?
Green = hit rate; pale band = uncertainty; blue bars = number of alerts.
Which past markets would have set it off?
Green ended Yes. Red is a false alarm. Orange line is your current threshold.
Confidence levels in this dataset
Why the sliders look “too good”
Loading…
Rows selected by the current score
Advanced / optional
Exact rule knobs behind the alarm
Skip this section on first read. It is for stress-testing the alarm. If a slider combination looks perfect but only selects a few related rows, it is not a reliable law of nature.
Resolved-market scatter
Orange points are rows selected by the current line.
Rule diagnostics: sample size and clustering
Loading…
Sensitivity around the current slider settings
Markets selected by the current rule
User-requested case
Ukraine invasion warning: how much heads up?
Forecast timeline before invasion
Community aggregate points vs high-intensity forecaster points; sourced reconstruction, not authenticated API raw dump.
Practical alert thresholds
Interpretation: an aggregate ≥50% was a “weeks ahead, take contingency planning seriously” signal; ≥60% on Feb 13 was a “leave Kyiv / prepare now” signal; expert/intensive monitoring reached 90%+ roughly 12 days out.
Ukraine evidence points
Pre-public timing check
Did markets move before public timestamps?
First crossing ≥80% relative to public source time
Bars are clipped to keep the chart readable; full values are in the table.
Timing rows
Examples and controls
Case studies
Auditability
Raw data and caveats
Local files used
Critical caveats
- Backtest windows are often anchored to Polymarket close/resolution; only selected rows have independent public-news anchors.
- Markets are not independent. Sibling markets around Iran/Venezuela can inflate confidence and p-values.
- Old Ukraine 2022 Polymarket CLOB data was not available through current endpoints; Ukraine section uses sourced public forecasting points.
- This is a monitoring/triage dashboard, not proof of insider trading and not trading advice.