Overview

This page displays the activation status of Anticipatory Action (AA) triggers at the national level in Zambia, as well as for ten key regions (referred to as districts here) during the Nov-Dec-Jan 2024/2025 season, based on forecasts issued in October.

For more details on the AA protocol and its implementation, click on the “AA Protocol” tab at the top of this dashboard.

National-Level Activation

Regional-Level Activation

Map Considerations

The Zambia map displays only the results of regional forecasts. For national-level results, refer to the activation section above.

Severe Scenario Table

Moderate Scenario Table

Mild Scenario Table

Accuracy Rate: the proportion of correct predictions (both positive and negative) made by the model out of all predictions.

This metric gives an overall sense of how often the model is right. However, it can be misleading in cases of class imbalance (e.g., when one class is much more frequent than the other).

• Range: 0 to 1
• Reference point: depends on class distribution
• Interpretation:
  • 0: no correct predictions
  • ≈ 0.5: model performs no better than random guessing
  • > 0.7: generally acceptable performance
  • 1: perfect prediction

Formula:

\[ \begin{align*} Accuracy = \frac{a + d}{a + b + c+ d} \end{align*} \]

Hit Rate (HR): measures the proportion of actual positive cases that the model correctly identifies.

It reflects the model’s ability to detect relevant events or instances—especially critical in contexts like medical diagnoses or weather alerts.

• Range: 0 to 1
• Ideal value: 1 (perfect detection)
• Interpretation:
  • 0: no actual positive cases were detected
  • < 0.5: low detection capability
  • ≈ 0.5: moderate detection, room for improvement
  • > 0.5: good detection performance
  • 1: all actual positive cases were correctly identified

Formula:

\[ \begin{align*} HR = \frac{a}{a + c} \end{align*} \]

False Alarm Ratio (FAR): proportion of Incorrect Positive Alerts

The False Alarm Ratio measures the proportion of predicted positive cases that were actually false—i.e., how many alerts were raised unnecessarily. A lower FAR indicates that the model’s positive predictions are more trustworthy.

• Range: 0 to 1
• Ideal value: 0 (no false alarms)
• Interpretation:
  • 0: no false alarms—perfect precision
  • < 0.3: excellent—most alerts are justified
  • ≈ 0.5: half of the alerts are incorrect—moderate reliability
  • > 0.5: unreliable—alerts are often wrong
  • 1: all alerts are false—no useful signal

Formula:

\[ \begin{align*} FAR = \frac{b}{a + b} \end{align*} \]

Bias Score (BS): Forecast Bias Indicator

The Bias Score quantifies a model’s systematic tendency to overpredict or underpredict events compared to actual observations. It doesn’t assess accuracy per se, but rather the balance of predictions.

• Range: 0 to ∞ (sometimes normalized around 1)
• Reference point: 1 (perfect balance)
• Interpretation:
  • < 1: underprediction – the model forecasts fewer events than actually occur
  • = 1: balanced prediction – the number of predicted and observed events match
  • > 1: overprediction – the model forecasts more events than actually occur
  • Example: a bias score of 0.7 suggests the model tends to miss important events, while a score of 1.2 indicates it tends to raise too many alerts

Formula:

\[ \begin{align*} BS = \frac{a + b}{a + c} \end{align*} \]

Hanssen-Kuipers Score (KSS): True Skill Statistic

The Hanssen-Kuipers Score measures a model’s ability to correctly discriminate between events and non-events, accounting for both correct predictions and errors.

• Range: –1 to 1
• Interprétation :
  • -1: predictions are completely wrong (inverse of reality)
  • 0: no skill—equivalent to random guessing
  • 0.3 – 0.5: useful model, but with room for improvement
  • 0.6 – 0.8: good discriminatory power
  • 1: perfect discrimination—no misclassification

This score evaluates both sensitivity (true positive rate) and specificity (true negative rate), making it especially robust in situations with imbalanced classes or rare events.

Formula:

\[ \begin{align*} KSS = HR - FAR = \frac{a}{a + c} - \frac{b}{b + d} = \frac{ad - bc}{(a + c)(b + d)} \end{align*} \]

Heidke Skill Score (HSS): Forecast Skill Relative to Random Chance

The Heidke Skill Score evaluates the overall accuracy of a prediction system by comparing its performance to that of a random forecast. Unlike simple accuracy, HSS accounts for the correct predictions that could occur purely by chance.

• Range: –∞ to 1 (typically –1 to 1 in practice)
• Interpretation:
  • < 0: worse than random guessing
  • 0: no skill—equivalent to chance
  • 0.3 – 0.5: useful model, but with room for improvement
  • 0.6 – 0.8: good predictive skill
  • 1: perfect classification—every instance correctly predicted

This score is especially valuable when dealing with imbalanced datasets, where accuracy alone can be misleading. HSS provides a more equitable assessment by factoring in the expected performance of a random model.

Formula:

\[ \begin{align*} HSS = \frac{2(ad - bc)}{(a + c)(c + d) + (a + b)(b + d)} \end{align*} \]

Operation of the Anticipatory Action Mechanism

The Anticipatory Action mechanism consists of a predictive element and an observational element.

Predictive triggers rely on seasonal rainfall forecasts for July-August-September, developed by IRI. If forecasts indicate a deficit season or the national cumulative precipitation average remains within the driest 30% of past seasons, the threshold is considered reached.

Evaluation Period

The evaluation period runs from January to August. Each month from January to June, seasonal forecasts are updated, and the trigger status is assessed. No assessment is conducted in July. In August, rainfall data from June and July is used to determine whether the observational trigger has been reached.

Predictive Trigger Activation

Evaluation of the predictive trigger activation (January to June, inclusive)

• Verification of threshold completion.
• Update of the Maproom tool.
• Production of the technical report.
• [Deadline: 22nd of the month] Communication of trigger activation status (see Section 6).

Figure 1. Timeline of the Triggering System for the NDJ Season in Zambia

Section to be determined—placeholder text follows.

Anticipatory Actions Triggered in Previous Months by Scenario

Please note that the current triggering system tracks national-level results, while regional data is provided for reference only. Below are the months in which each scenario has been triggered so far.

AA Package for Triggering

Section to be determined