Acestor Documentation

Acestor is a production dengue-intelligence pipeline. It ingests case and weather data, estimates per-region risk thresholds, runs an ensemble of forecasting models (RF, XGB, TSE, NBR, TimesFM), produces interactive risk maps, and emits a self-contained HTML brief — all driven from a single YAML configuration file.

Get Started

New here? Run your first pipeline in a few minutes.

Quickstart

GitHub

Browse the source code, open issues, and contribute.

https://github.com/dsih-artpark/acestor

---

Three Pipelines

Acestor is not a single monolithic pipeline — three independently-runnable DAGs share the same config surface:

Pipeline	What it does	Config example
dengue_prep	Ingest raw case data (IHIP linelist / dashboard API / custom source) + weather (Open-Meteo / cache / custom) → produce cleaned prepared_data/*/cases_daily.csv and weather_daily.csv	configs/ka_district_prep.yaml
dengue	Load prepared data → compute thresholds → train models → write predictions.csv + interactive HTML brief	configs/ka_district.yaml
dengue_downscale	Take a parent-level predictions.csv and apportion it to child regions via observed case shares + Largest Remainder Method (integer conservation guarantee)	configs/ka_district_to_subdistrict.yaml

Main-Pipeline Steps

The dengue pipeline is a 11-step DAG — steps run concurrently where dependencies allow.

#	Step	What it does
1	identify_sampling_day	Derives the weekly anchor (W-MON/W-SAT/…) from run.run_date
2	parse_case_data	Loads and re-samples prepared_data/{region_type}/cases_daily.csv to the sampling grid
3	parse_weather_data	Loads weather_daily.csv, rolls to weekly with configurable lags
4	validate_case_data_sufficiency	Early gate — stops if data is too thin (min rows / regions / span)
5	identify_cutoff_dates	Picks case + weather cutoffs and the 4-week prediction calendar
6	generate_thresholds	Per-region historical / prev_nweeks / weighted_baseline threshold tables
7	train_and_predict	Fits every model in model.models (RF, XGB, TSE, NBR, TimesFM), writes per-model + ensemble predictions
8	assess_thresholds	Threshold assessment table + figure metadata
9	generate_maps	Static PNG choropleth maps (skip via maps.enabled: false — HTML uses interactive D3 maps regardless)
10	generate_report	Single-file report.html with hero chart, weekly zone tables, interactive D3 map
11	send_report	Emails the report (opt-in via email.enabled: true)

The dengue_prep and dengue_downscale DAGs are documented in dengue_prep Pipeline and Downscale Pipeline.

Key Features

• Pluggable data sources. Case data from local files, the dengue-dashboard API, or a custom Python module. Weather from Open-Meteo, filesystem cache, or a state API. See Case & Weather Source Plugins.

• TimesFM 2.5 forecasting. Google’s pretrained time-series foundation model, run in an isolated subprocess so torch can’t collide with xgboost’s OpenMP.

• Integer-consistent downscaling. Largest Remainder Method guarantees sum(child.predictionInt) == parent.predictionInt.

• Historical-window fallback. When the recent 4-week share window is empty, fall back to a 52-week window before uniform-splitting.

• Percentile risk classification. Per-region historical percentile cutoffs (default 50/75/90) — plus WHO and ICMR classifiers available in parallel.

• Interactive HTML brief. D3-based choropleth + zone tables in a single report.html, no server required.