Skip to model research
Systematic Trading ModelsSpecification / Evaluation / Release

Probability-group and threshold lens

Constant declared probabilities against authored binary labels. Forecast loss and threshold decisions remain separate.

Exact only for constant probabilities

Every observation in a declared group receives exactly its entered probability. A group mean from varying row forecasts is insufficient for this exact binary loss. Counts must be whole non-negative values, with positives no greater than observations. Empty groups have undefined observed fractions and gaps but zero weight. The combined population must contain at least one row. Inputs are declarations, not authenticated market data. The default one hundred labels are authored for education. The downloadable specimen expands the entire default without inventing an instrument, chronological source, learned model or real probability estimate.

Binary loss and separate decisions

Brier loss adds squared probability error for every positive and negative label and divides by total observations on the binary zero-one scale. The threshold is not part of that loss. It uses probability greater than or equal to threshold to form positive decisions; equality is included. Confusion metrics reuse the original count worksheet's definitions. A zero denominator stays undefined. Changing the threshold changes selected counts, not supplied forecasts. The exercise cannot reconstruct varying probabilities from a confusion matrix, choose an optimal trading action or demonstrate an economic result from a label score.

The reference is fitted to this same table

The prevalence reference probability is the entered positive-label fraction, q. Its binary loss is q times one minus q. Relative comparison is one minus entered loss divided by that reference loss; it is undefined when the denominator is zero. This is not a benchmark frozen before the table's outcomes. Its name remains same-table throughout the display and export. A favorable ratio has no return units and is not financial alpha. A proper real comparison needs a predeclared reference policy, population, probability generation and reserved evaluation rather than a retrospective table-only claim.

A group gap is descriptive, not a fitted calibration model

Each group's gap compares its constant probability with its observed label fraction. The weighted absolute group gap combines those declared groups using observation counts. It does not learn bin edges, fit a calibration map or estimate future uncertainty. Zero gap in the default is deliberately authored. Binary Brier loss does not isolate calibration alone. A real forecast assessment needs genuine frozen predictions and labels under their source and selection contract. The site does not derive numerical win probabilities from conviction grades or transfer fictional matching frequencies to an actual provider.

Exports retain the absence of real evidence

The JSON keeps input declarations, exact calculation results, the constant-group requirement, binary scale and threshold operator. Actual model identity, forecasts and financial outcomes remain null. Invalid edits clear the previous result and disable export. Reset and reload return to the authored default. There is no account connection, data upload, local-storage persistence or automated trading action in this lens. The dedicated public search is a separate URL-query feature. Do not put secrets or customer records into public browser forms or treat a downloaded assumption record as authenticated research evidence.

Economic validation still needs its own process

The tool omits payoffs, exits, quantity, weights, fees, financing and actual execution. Class labels are not defined as profitable trades. Known reference checks establish software correctness only. A genuine instrument research task must complete the authoritative factory and independently verified master, preserving raw causal entry ranking, selected native exit and fill convention. A website can publish useful model questions without claiming that it has trained or certified a financial candidate. No new model promotion, cloud pipeline, production scheduler or paid capacity is authorized by this exercise.