← ai·maturity Lab Demo · research model snapshot · runs in your browser
Worum geht es

Das Konfliktmodell liefert Wahrscheinlichkeiten, aber ab welcher Schwelle soll ein Analyst eine Akte öffnen? Verschieben Sie die Schwelle, setzen Sie Teamkapazität und die Kosten von Fehlalarm und Verpasstem. Erwartete Treffer und Fehlalarme folgen direkt aus den kalibrierten Wahrscheinlichkeiten. Achten Sie auf die niedrige Basisrate.

What this is about

The conflict model yields probabilities, but above which threshold should an analyst open a dossier? Move the threshold, set team capacity and the cost of a false alarm versus a miss. Expected hits and false alarms follow directly from the calibrated probabilities. Note the low base rate.

Where do you set the alarm threshold? loading model

Early warning on a calibrated conflict-forecast model (model snapshot –, – countries). A model only produces probabilities; someone still has to decide above which probability an analyst opens a dossier. Move the threshold, set the team's capacity and the relative cost of a false alarm versus a miss — every figure is an expectation under the model's own calibrated probabilities, recomputed instantly.

Flagged
–
dossiers to watch
Expected true positives
–
Σ p over flagged
Expected false alarms
–
Σ (1 − p) over flagged
Expected misses
–
Σ p over unflagged
Expected precision · recall
–
share of flags that are real · share of events caught
Expected cost
–
relative units

Expected precision–recall one point per distinct probability level

model, all thresholdscurrent threshold

Expected cost vs threshold log scale

totalfalse-alarm costmiss costcost-optimal t*

Base-rate strip one dot per country, sorted by probability

–

flagged (p ≥ t)not flagged, p > 0p = 0 in this calibration
hover a dot for the country

Flagged countries click a column header to sort

ISO3 Country Status p E[false alarm] Deaths last month Months since active Dominant type

How it works

The snapshot (–) gives every country a calibrated probability for the chosen target and horizon: for currently peaceful countries the probability of a conflict onset (the early-warning case), for all countries the probability of active conflict. An alarm rule is simply a threshold t: the flagged set is F = {countries with p ≥ t}. Because the probabilities are calibrated, the consequences of a rule can be estimated before anything happens and without a ground truth: expected true positives ΣF p, expected false alarms ΣF (1 − p), expected misses Σnot F p, expected precision ΣF p / |F|, expected recall ΣF p / Σall p, and expected cost cFA · E[false alarms] + cmiss · E[misses]. Flagging a country pays off when its expected miss cost exceeds its expected false-alarm cost, i.e. when p > cFA / (cFA + cmiss) — that ratio is the cost-optimal threshold t*, independent of the model. The capacity rule instead takes the k-th highest probability as threshold (ties are flagged together, so the list may exceed k). The naive rule for comparison flags, for the early-warning target, every peaceful country that recorded fatalities in the last month; for the active-conflict target it flags every country that is active today (status quo). Its expected outcomes are evaluated with the same model probabilities.

Read the numbers as expectations, not outcomes. They are what the model believes will happen on average if its probabilities are right. If the model is miscalibrated — say its "30 %" countries actually experience onsets 10 % of the time — then every expectation here is wrong by the same amount. That is why the Brier score is shown next to the AUC: a model can rank countries almost perfectly (AUC 0.9) and still be useless for threshold-setting if its probability scale is off. The calibrated probabilities in this snapshot are visibly binned (only a handful of distinct levels per horizon), which is why the precision–recall curve is a short staircase rather than a smooth curve.

What this teaches about AI maturity

A threshold is not a property of the model; it is a business decision that encodes two things the model cannot know — how many dossiers the organisation can actually work, and how much a false alarm costs relative to a miss. Mature organisations make that trade-off explicit and revisit it, instead of accepting a default of 50 %. They also look at base rates first: with roughly one expected onset among 155 peaceful countries per year, even an excellent ranker produces mostly false alarms, and "precision" has to be judged against that base rate, not against intuition. Finally, they insist on calibration (Brier, reliability curves) rather than celebrating AUC alone, because every downstream decision runs on the probability scale.

Research and teaching demo, not policy, security or investment advice. The risk figures are a frozen snapshot of a research model (model state –), not a live feed; the model's historical descriptions are not shown. Country names are ISO-3 codes with English short names. Expected values are computed under the model's own calibrated probabilities and inherit any miscalibration. Everything runs in your browser; nothing is stored or transmitted.