Output Format

Every pipeline run writes its artifacts under {storages.artifacts.filesystem.base_path}/{run_id}/. The canonical prediction table is always at outputs/predictions.csv regardless of the model / output mode chosen. This page documents every column in that CSV, the auxiliary outputs (charts, maps, interactive HTML brief), and the Largest-Remainder invariant that ties parent and child predictions together in the downscale pipeline.

Directory layout

Path (relative to run root)	What it is
inputs/cases_<res>_sampled.csv	Case data trimmed to the modelling window
inputs/weather_<res>_sampled.csv	Weather features aligned to the modelling window
inputs/thresholds_<res>.csv	Long-form (region_id, date, Mean, StdDev, threshold_method) table covering every configured method, used as input to train_and_predict
outputs/predictions.csv	Canonical predictions table (see column reference below)
outputs/per_model/predictions_<model>.csv	Per-model copies (only in output: both or output: per_model modes)
outputs/charts/hero_forecast.png	Region-level forecast time-series chart embedded in the HTML brief
outputs/maps/*.png	Choropleth PNGs, one per (model, thresholdMethod, week) — only generated when maps.enabled: true
outputs/report.html	Interactive HTML brief with an embedded D3 choropleth (see The HTML brief below)
outputs/prep_summary.md	Human-readable summary of the downscale step (downscale runs only)

predictions.csv — column reference

A real header line from artifacts/ap/ap-repro-fixed/outputs/predictions.csv:

recordDate,thresholdMethod,Mean,StdDev,Zero,Inf,T0.00,T1.00,T2.00,
startDatePredictedWeek,dateOfComputingPrediction,regionID,lgdCode,
predictionRaw,ISOWeek,model,predictionMin,predictionMax,
predictionZone,whoZone,prediction,icmrZone,percentileZone

A real data row from the same file:

2026-03-17,previousNweeks,1.6666666666666667,0.478713553878169,
0.0,inf,1.6666666666666667,2.1453812205448357,2.6240957744230045,
2026-03-17,2026-07-24,district_502,502,1.5658082125731811,12,
ensembleModel,,,3,1.0,2,3,2

Columns are grouped by role below.

Identity columns

Column	Meaning
regionID	Fully qualified region id (e.g. district_502, ward_198, zone_5). Always prefixed with the spatial resolution.
startDatePredictedWeek	ISO date (Mon-anchored) of the week the prediction is for.
dateOfComputingPrediction	Wall-clock date the run executed. Useful for hindcast reproducibility audits.
ISOWeek	ISO week number (1..53) of startDatePredictedWeek.
model	Model that produced the row: nbr, timesfm, or ensembleModel. In output: ensemble mode this is always ensembleModel.
thresholdMethod	Threshold-generation method whose Mean/StdDev feed this row’s T* columns. Values: historical, previousNweeks, weightedBaseline (camelCase per PRISM-H).
lgdCode	Local Government Directory numeric code, added by pipelines/dengue/lib/lgd.py (PR #59). Blank if the region isn’t in the LGD map.

Numeric prediction columns

Column	Meaning
prediction	Display integer. This is the value dashboards and briefs show. Rounded from predictionRaw with round_half_up (0.5 rounds up), clamped ≥ 0. In the downscale pipeline it also satisfies the LRM invariant (see LRM invariant).
predictionRaw	Raw float from the model / ensemble. Used by backtesting, calibration, and downscale shares. Never rounded internally — the CSV boundary is the only place the rename happens (train_and_predict.py: prediction ⇄ predictionRaw).
predictionMin, predictionMax	Inter-ensemble range (PR #87 / issue #84). For ensemble rows, these are the min and max of the corresponding per-model predictions for the same (region, date). For per-model rows, min == max == prediction. Not a confidence interval; not std; not bootstrap. Semantic: extremes across the members. The CSV also enforces predictionMin ≤ prediction ≤ predictionMax at the boundary, widening the bracket if rounding pushed the display int outside.
predictionInt (downscale only, internal)	LRM-consistent integer produced by largest_remainder in the downscale step. Renamed to prediction at CSV write, so it does not appear in the header. See LRM invariant.

Threshold columns

The threshold columns are copied from the inputs/thresholds_<res>.csv table for the (regionID, startDatePredictedWeek, thresholdMethod) key.

Column	Meaning
Mean	The μ value for the row’s thresholdMethod. What “recent” or “historical” or “weighted baseline” is depends on the method (see docs/threshold-methods-reference.md).
StdDev	The σ value. Inflated to sqrt(Mean) when it collapses to 0 while Mean > 0 (PRISM-H §4.4).
T0.00, T1.00, T2.00	Threshold cut-points at α ∈ list_alpha. Formula: T_α = Mean + α · StdDev + i · 1e-6 (i = index of α in the list, 1-based, ensuring strict monotonicity when Mean/StdDev collapse). Default list_alpha: [1.0, 2.0] produces three T-columns.
Zero, Inf	Sentinel bounds (always 0.0 and inf). Combined with the T-columns to form [Zero, T0.00), [T0.00, T1.00), …, [T2.00, Inf) half-open intervals that predictionRaw is bucketed into to derive whoZone.

Zone assignment columns

Three parallel classifications are always computed and always written. predictionZone is a copy of one of them, selected by thresholds.classification_method in the config.

Column	Meaning
whoZone	WHO-style band. Integer in 1..N where N = len(list_alpha) + 2. Derived by binning predictionRaw into the [Zero, T0.00, T1.00, …, Inf] intervals. Set to NA when the row’s Mean and StdDev are both zero (degenerate threshold).
icmrZone	ICMR quartile stratum (A1–A4 → 4..1). Cross-sectional per date: distinct predicted values across all regions on that date are ranked descending and cut into four equal strata. Set to NA when the total predicted caseload for the date is < 10 (PRISM-H §5.4 “insufficient data” guard).
percentileZone	Per-region historical-percentile band (PR #73, issue #62). Cut points come from the region’s own weekly-aggregated case history at percentile_cutoffs (default [50, 75, 90] → 4 bands: ≤ p50, (p50, p75], (p75, p90], > p90). Regions with no history get NA.
predictionZone	Convenience copy of whichever of the above is selected by thresholds.classification_method (who / icmr / percentile). NA values are filled with the sentinel 0 — dashboards render zone 0 as light grey and log a warning listing the affected regions.

LRM invariant

The downscale pipeline (pipelines/dengue_downscale) apportions each parent-level predictionRaw across its children using the Largest Remainder Method (Hamilton’s method, apportionment.py).

For every (parent, week):

sum(child.prediction) == round_half_up(sum(child.predictionRaw))
                      == round_half_up(parent.predictionRaw)

Each child’s integer allocation is either floor(child_raw) or floor(child_raw) + 1 — off by at most one from the natural round. Ties on fractional remainder are broken by (a) higher recent-cases wins (issue #86), then (b) lower region_id lexicographically. The downscale step logs sum(predictionInt) vs round_half_up(sum(prediction)): M/N parent-weeks match — this should always be N/N.

Charts and maps

• outputs/charts/hero_forecast.png — the region-focused forecast chart embedded at the top of report.html. Always produced.

• outputs/maps/<model>_<thresholdMethod>_<YYYYMMDD>.png — a static choropleth per prediction week. Only produced when maps.enabled: true in the config. The set is deterministic: one image per (model in outputs, thresholdMethod in outputs, startDatePredictedWeek) combination.

The HTML brief

outputs/report.html is a fully self-contained HTML document — no external assets. It embeds:

• The hero chart as an inline PNG.

• A per-week narrative summary.

• An interactive D3 choropleth. All map data is inlined as a JS global by pipelines/dengue/templates/partials/_d3_map_script.html.j2:

  window.__BRIEF_DATA = {
    geojson:       <FeatureCollection of all rendered regions>,
    parent_lookup: {<parent_id>: {name, children: [...], ...}},
    weekly_zones:  {<week_index>: {<region_id>: <zone_int>}}
  };
  

  geojson is the union of every parent polygon that has predictions in the brief; parent_lookup powers the drill-in / drill-out behaviour; weekly_zones is the per-week zone assignment used to paint the polygons. All three are trimmed to the horizon that the brief covers, keeping report.html small enough to email.

Output modes

model.output controls what ends up on disk:

Mode	Canonical outputs/predictions.csv	outputs/per_model/predictions_<m>.csv
ensemble (default)	Ensemble rows	(not written)
both	Ensemble rows	One file per non-ensemble model
per_model	Copy of report.primary’s per-model file	One file per configured model

The canonical path never changes — consumers (officials, the HTML brief, the downscale pipeline) can always read outputs/predictions.csv without needing to know the mode.