Expected return10% trimmed meanHistorical drift
Market brief
01
Expected return path
The forecast is anchored on the mean terminal return. A robust trimmed-mean cross-check is shown without turning tail outcomes into the visual headline.
02
Probability × payoff
Why a 65–75% directional hit rate can still coexist with meaningful downside when the thesis fails.
03
Two independent lenses
Nearest full-stack historical structures are checked against an expectation implied directly by the current boundary and moderate event probabilities.
04
Targets that matter now
Boundary and moderate forecasts remain visible; the current decision filter determines which ones can drive the summary.
05
Recent thesis journey
Hover or select any issue date to see which target calls resolved correctly, which missed, and which remain active.
Forecast map
01
Core decision surface
Boundary and moderate forecasts define the main decision surface. The active confidence floor is shared with Market Brief.
Probability that the target is reached at any point inside the horizon.
Shrunk historical accuracy for the same target, selected class, and confidence band.
Neutral below 55% selected confidence; supported at 55–60%; decision-grade at 60%+.
02
Tail watch
Tail events solve a different problem. Instead of forcing a 50% Event threshold, compare each tail probability with its own historical occurrence rate and probability distribution.
03
All 18 forecast calls
The complete surface remains visible and auditable. Core calls emphasize confidence-conditioned reliability; tails emphasize event probability versus their normal base rate.
| Window | Target | Model view | Event prob. | Selected conf. | Reliability / tail context | Edge vs base | Call holds | Call breaks | History |
|---|
04
What is driving the thesis
Only non-tail forecasts outside the neutral zone contribute materially to the directional synthesis.
Model evidence
Show the work behind the market brief.
Probability skill, base-rate-adjusted reliability, return robustness, analog quality, sample stability, and forecast-level diagnostics.
01
Confidence → reliability frontier
Higher selected-class confidence improves reliability and probability skill, but at the cost of fewer decision-ready forecasts. This is the explicit coverage tradeoff behind the neutral zone.
02
Precision–recall explorer
Every horizon × target has its own discrimination curve. Tails often rank rare events well even when they almost never cross a 50% Event threshold.
03
Return robustness and baseline
Expected return remains the full arithmetic mean. Median and trimmed mean expose skew; historical drift shows how much incremental edge the current structure adds.
| Window | Expected | Middle 50% | Baseline | Incremental edge | Median | 10% trimmed | Non-overlap blocks |
|---|
04
Lens reconciliation
Analog outcomes versus probability-implied terminal returns.
05
Integrity checks
What could make the current read fragile.
06
Forecast-level evidence
“Prob. skill” is Brier Skill Score versus a constant target base-rate forecast. It evaluates the probabilities themselves, not merely the 0/1 class decision.
| Window | Target | Event prob. | Selected conf. | Decision tier | Reliability at confidence | Base rate / tail base | Edge | Prob. skill | Recent skill | Return implied |
|---|
07
Conditional outcome diagnostics
Detailed return and path behavior stays here so it can inform trust without dominating the executive story.
| Window | Target | Call | If call holds | If call breaks | Median hold / break | Failed occurrence | Return population |
|---|
Metric definitions and guardrails
Reliability edge = selected-class historical accuracy minus the unconditional accuracy of that same class. It prevents easy non-occurrence calls from looking strong merely because the event is rare.
Accuracy skill rescales reliability edge by the remaining distance from the base rate to 100%. A 20% score means the model closes 20% of the possible classification improvement beyond the base rate.
Brier skill compares squared probability error with a constant historical event-rate forecast. Positive is better than climatology; zero is no probabilistic improvement.
Stack-implied return combines each current boundary/moderate event probability with historical terminal returns conditional on target hit versus no hit, then takes the median across the four primary targets in a horizon. It is a cross-check, not the headline expected return.
History lab
Interrogate the forecast archive.
Filter actualized historical forecasts to understand reliability, calibration, miss paths and return distributions under specific conditions.
Source
Share of selected-class forecasts that resolved correctly.
Median share of the target realized when it was not reached.
Median furthest adverse move among those same target misses.
Rolling reliability
Rolling selected-class accuracy by forecast issue date.
Calibration
Predicted event probability versus realized event frequency.
Target-miss profile
Typical target realization and adverse excursion when a target was not reached.
Reliability by horizon and threshold
Which forecast families have historically been most dependable under the active filters.
Terminal return distribution
Final market return across the filtered forecast set.
Time to hit
How quickly target events were reached when they occurred.
Forecast detail
Most recent realized rows under the active filters.
| Date | Window | Target | Selected class | Prediction | Result | Reliability | Return | Agreement |
|---|