Province forecasts

This page gives the one-week-ahead forecast for each province from the joint model. The projection is defined in the province forecast Methods section. The national forecast is on the forecasts page. How these forecasts have scored against what each province went on to report is in the forecast by province evaluation. How well the model reproduces each province's share is on the in-sample checks page.

Load packages, data and fitted chains

# Shared setup: packages, observations and the fit registry. See
# `docs/pages/_setup.jl`.
using BVDOutbreakSize
include(joinpath(pkgdir(BVDOutbreakSize), "docs", "pages", "_setup.jl"))

validation_forecast_from (generic function with 1 method)

# The fits this page reads, loaded from the cache here.
chn_joint = load_fit("joint");

Summary

The table gives each province's forecast for the week to   as a 90% predictive interval, and the bullets under it compare the provinces draw by draw. The provinces add up to the national forecast.

Project each province a week ahead

# Read from the same draws as the national forecast page, so the provinces
# add up to the national forecast shown there.
province_draws = fit_forecast("joint");
province_projection = forecast_provinces(
    province_draws; horizon = 7, n_patches = N_PATCHES
);
province_forecast = province_forecast_table(
    province_draws, province_projection;
    n_patches = N_PATCHES
);
province_forecast_fig = plot_province_forecast(
    province_draws, province_projection;
    n_patches = N_PATCHES
);
province_week_end = obs.cutoff + Day(7);
province_forecast_md = province_forecast_headline(
    province_projection; n_patches = N_PATCHES
);

Province	New confirmed cases	New confirmed deaths	Patients in isolation	Isolation beds	New infections	R at T+7	P(R > 1)
Ituri	159–394	66–196	259–653	959–1388	169–684	0.49–1.12	13%
Nord-Kivu	68–241	47–163	104–362	242–353	90–494	0.44–1.23	18%
Haut-Uele	9–71	2–35	7–71	131–194	16–140	0.46–1.57	35%
Other provinces	2–29	0–14	4–57	20–31	7–66	0.57–1.75	52%

• Most new confirmed cases: Ituri, with probability 92%.

• Most new confirmed deaths: Ituri, with probability 70%.

• Growing at T+7: 1 of 4 provinces are more likely than not to be growing (Other provinces).

• Beds full at T+7: no province is more likely than not to be at or above its beds.

One-week-ahead forecast by province

Each panel is one forecast target, the provinces side by side, for the week to province_week_end.

The maps shade each province by its patch's forecast, so the pooled patch shades all its provinces alike. The reproduction number is centred on one, and a province whose 90% predictive interval spans one is washed out.

Build the forecast maps

province_forecast_draws(col) = [
    float.(province_projection[province_projection.patch .== p, col])
        for p in 1:N_PATCHES
]
province_forecast_map = plot_province_map(
    filter(
        !isnothing,
        [
            (;
                province_map_summary(province_forecast_draws(:rt_forecast))...,
                title = "Reproduction number at T+7", diverging_at = 1.0,
                colorbar_label = "R (median)",
            ),
            (;
                values = province_map_summary(
                    province_forecast_draws(:confirmed_new)
                ).values,
                title = "New confirmed cases by T+7",
                scale = CairoMakie.Makie.pseudolog10,
                colorbar_label = "Confirmed cases (median)",
            ),
            :isolation_level in propertynames(province_projection) ? (;
                    values = province_map_summary(
                        province_forecast_draws(:isolation_level)
                    ).values,
                    title = "Patients in isolation at T+7",
                    scale = CairoMakie.Makie.pseudolog10,
                    colorbar_label = "Patients (median)",
                ) : nothing,
        ]
    )
);

Forecast summary table by province

Province	Quantity	Lower 90%	Lower 60%	Lower 30%	Upper 30%	Upper 60%	Upper 90%
Ituri	New confirmed cases by T+7	159	205	232	289	324	394
Ituri	New confirmed deaths by T+7	66	93	110	140	160	196
Ituri	Patients in isolation at T+7	259	332	380	469	525	653
Ituri	Isolation beds at T+7	959	1050	1102	1203	1269	1388
Ituri	New infections by T+7	169	246	303	416	497	684
Ituri	Reproduction number at T+7	0.49	0.63	0.71	0.86	0.94	1.12
Nord-Kivu	New confirmed cases by T+7	68	99	118	156	183	241
Nord-Kivu	New confirmed deaths by T+7	47	67	81	107	124	163
Nord-Kivu	Patients in isolation at T+7	104	151	178	237	279	362
Nord-Kivu	Isolation beds at T+7	242	265	278	305	322	353
Nord-Kivu	New infections by T+7	90	135	171	253	325	494
Nord-Kivu	Reproduction number at T+7	0.44	0.57	0.66	0.85	0.98	1.23
Haut-Uele	New confirmed cases by T+7	9	18	24	38	49	71
Haut-Uele	New confirmed deaths by T+7	2	5	8	15	21	35
Haut-Uele	Patients in isolation at T+7	7	14	20	34	44	71
Haut-Uele	Isolation beds at T+7	131	143	151	166	175	194
Haut-Uele	New infections by T+7	16	28	37	63	83	140
Haut-Uele	Reproduction number at T+7	0.46	0.63	0.75	1	1.19	1.57
Other provinces	New confirmed cases by T+7	2	5	7	14	18	29
Other provinces	New confirmed deaths by T+7	0	1	2	5	7	14
Other provinces	Patients in isolation at T+7	4	10	15	27	36	57
Other provinces	Isolation beds at T+7	20	22	23	26	28	31
Other provinces	New infections by T+7	7	12	16	27	36	66
Other provinces	Reproduction number at T+7	0.57	0.76	0.9	1.14	1.32	1.75

Forecast for each province

Each province has its own forecast figure, one histogram per target with its 90% predictive interval shaded. The dashed rule on the case and death panels is the count the province reported over its most recent week in the spatial tables.

Build the per-province figures

# The spatial tables' most recent week in each province, clamped as the
# compositions are, so a downward revision reads as no new cases.
recent_cases = province_recent_counts(
    obs.province_confirmed_history, PROVINCE_NAMES, N_PATCHES
)
recent_deaths = province_recent_counts(
    obs.province_death_history, PROVINCE_NAMES, N_PATCHES
)
function forecast_province_observed(p)
    o = (;)
    recent_cases === nothing ||
        (o = merge(o, (; confirmed_new = recent_cases.counts[p])))
    recent_deaths === nothing ||
        (o = merge(o, (; confirmed_deaths_new = recent_deaths.counts[p])))
    return o
end
forecast_province_fig(p) = plot_province_forecast_detail(
    province_draws, province_projection;
    province = p, n_patches = N_PATCHES,
    observed = forecast_province_observed(p)
)
# The province blocks below are written out one per patch.
@assert N_PATCHES == 4 && PROVINCE_LABELS[1:4] ==
    ["Ituri", "Nord-Kivu", "Haut-Uele", "Other provinces"]
# Cases and deaths come from separate tables, so each stream's observed
# week is stated on its own.
function forecast_province_recent_week(r, stream)
    r === nothing &&
        return "The spatial tables carry no recent week of $(stream)."
    return string(
        "The observed week of $(stream) runs from ",
        grid_date(r.start_day), " to ", grid_date(r.last_day), "."
    )
end
recent_note = join(
    [
        forecast_province_recent_week(recent_cases, "confirmed cases"),
        forecast_province_recent_week(recent_deaths, "confirmed deaths"),
        "Each ends at its last spatial vintage, which can be earlier than " *
            "the cut-off.",
    ], " "
);

The observed week of confirmed cases runs from 2026-09-20 to 2026-09-27. The observed week of confirmed deaths runs from 2026-09-20 to 2026-09-27. Each ends at its last spatial vintage, which can be earlier than the cut-off.

Ituri

Nord-Kivu

Haut-Uele

Other provinces

The other provinces are pooled into one patch, so their forecast is for the pool.

Past forecasts against what was observed

Only province forecast projections are shown, so the figure fills in as releases accumulate. Each panel is one province stream at one horizon. The x-axis is the cut-off each forecast was made from. Each forecast shows its median and 90% predictive interval, beside the persistence baseline and the count the province went on to report. A window holding a harmonisation-break day is left out, because that day's backfill is published for the country and not by province.

Load the archived province forecasts and their outcomes

# Written by `scripts/score_releases.jl` from each release's
# `province_forecast.csv`, in the national overlay's schema. A missing
# file reads as an empty table, which the figure reports as nothing scored.
province_overlay_df = _release_data(
    joinpath("province", "forecast_overlay.csv"),
    (;
        release = String, made_date = Date, stream = String, horizon = Int,
        target_date = Date, fit = String, observed = Float64,
        median = Float64, lo30 = Float64, hi30 = Float64, lo60 = Float64,
        hi60 = Float64, lo90 = Float64, hi90 = Float64,
    )
)
province_overlay_fig = plot_forecast_overlay(
    scored_overlay(province_overlay_df);
    empty_message = "No projection forecast has been scored yet. " *
        "This fills in as releases accumulate."
);

Saving province map assets

The dashboard map's province layer and its detail column read the per-province and national estimates and time series written here.

Write the province map estimates and time series

# The national row reads the joint's national reproduction number and the
# provinces' summed forecast.
province_map_dir = joinpath(
    pkgdir(BVDOutbreakSize), "docs", "src", "summary_assets"
)
mkpath(province_map_dir)
CSV.write(
    joinpath(province_map_dir, "province_estimates.csv"),
    province_map_estimates(
        chn_joint, province_forecast_draws(:confirmed_new);
        n_patches = N_PATCHES,
        confirmed_history = obs.province_confirmed_history,
        death_history = obs.province_death_history, cutoff = obs.cutoff
    )
);
national_forecast_draws = let d = province_projection
    [float(sum(d.confirmed_new[d.draw .== i])) for i in sort(unique(d.draw))]
end
CSV.write(
    joinpath(province_map_dir, "national_estimates.csv"),
    province_map_estimates(
        [vec(Array(chn_joint[:R_T]))], [national_forecast_draws];
        confirmed_history = Dict("national" => obs.confirmed_history),
        death_history = Dict("national" => obs.confirmed_deaths_history),
        cutoff = obs.cutoff, patch_names = ["national"],
        patch_labels = ["National"],
        members = Dict("national" => ["national"])
    )
);
# The daily reproduction number and the weekly confirmed cases of each
# patch and of the country, keyed by the patch label.
province_ts_rt_args = (;
    n = obs.n, breakpoint = _BREAKPOINT, rt_start = _rt_start_plot,
    rt_walk_start = clamp(_BREAKPOINT - RT_WALK_LEAD, _rt_start_plot, obs.n),
    ramp = RT_INTERVENTION_RAMP,
)
province_ts_inc = province_increment_matrix(
    obs.province_confirmed_history, PROVINCE_NAMES, N_PATCHES
)
national_ts_inc = let c = obs.confirmed_history.counts
    isempty(c) ? zeros(Int, 1, 0) : reshape([c[1]; max.(diff(c), 0)], 1, :)
end
# Tag a long table with the series it holds.
_with_series(t, s) = (t[!, :series] .= s; t)
CSV.write(
    joinpath(province_map_dir, "province_timeseries.csv"),
    vcat(
        _with_series(
            rt_quantile_table(
                [
                    reconstruct_patch_rt(
                        chn_joint; n_patches = N_PATCHES,
                        province_ts_rt_args...
                    );
                    [reconstruct_rt(chn_joint; province_ts_rt_args...)]
                ],
                [PROVINCE_LABELS[1:N_PATCHES]; "National"];
                cutoff = obs.cutoff, n = obs.n, from = _rt_start_plot
            ), "rt"
        ),
        _with_series(
            vcat(
                weekly_count_table(
                    province_ts_inc.days, province_ts_inc.increments,
                    PROVINCE_LABELS[1:N_PATCHES];
                    cutoff = obs.cutoff, n = obs.n
                ),
                weekly_count_table(
                    obs.confirmed_history.days, national_ts_inc,
                    ["National"]; cutoff = obs.cutoff, n = obs.n
                )
            ), "cases"
        );
        cols = :union
    )
);

---

The full analysis code, data and model definitions are in the epiforecasts/BVDOutbreakSize repository.