Run a forecast

Four steps: pick a series, look at the data, choose a method (one is recommended from its backtest record), then see the coming weeks and how far to trust them. Every number comes from the real data read at build time.

① Choose the series to forecast

Choose the series to forecast US national weighted ILI (wILI, %)United States · national · weekly · 835 weeks · CDCUS national ILI outpatient visitsUnited States · national · weekly · 835 weeks · CDCEngland: influenza test positivity (%)England · weekly · 483 weeks · UKHSAEU/EEA influenza positivity, sentinel (%)EU/EEA (aggregate) · weekly · 225 weeks · ECDCMalta: influenza detections, sentinelMalta · weekly · 51 weeks · ECDC

US national weighted ILI (wILI, %)

Decision Forecast allowed

② Data overview

Time span: Oct 3, 2010 – Sep 27, 2026
Frequency: One value per week
Length: 835 weeks (≈ 16.1 season(s))
Latest value: 1.85% (Sep 27, 2026)
Last 52 weeks: 0.87% – 8.3%
Missing: 0 weeks (0%), longest gap 0 weeks
Seasonality: Clear: strength 0.81 (cut-off 0.35)

Download this series JSON Parquet

History
Forecast
80% band / 95% band

The chart shows the last 104 weeks; the blank area on the right is the future to forecast.

Data check-up

No issues found.

③ Choose a forecasting method

Each method was replayed at 52 past points without seeing what came next, forecasting 8 weeks ahead each time (rolling backtest). Shorter bar = smaller average error.

Choose a forecasting method Theta methodRecommended Removes the season, smooths the trend, then adds the season back; parameters chosen automatically. Mean error 0.41 pp · 42% below baseline Lowest backtest error: 42% below repeating last year. MSTL decomposition Splits the series into seasonal and trend parts and extends each. Mean error 0.46 pp · 35% below baseline Error 12% higher than the recommended method. Naive Carries the latest week forward. Mean error 0.86 pp · 21% above baseline Did not beat the baseline (1.21× its error); not used on its own. Seasonal naive (baseline) Repeats the same week last year; every other method has to beat it. Mean error 0.71 pp Baseline, for comparison only.

Unavailable methods (1)

Seasonal ETS (exponential smoothing)Unavailable Season length is 52 weeks: statsforecast cannot fit a seasonal ETS that long (it collapses to Naive), so it is not a candidate.

④ Forecast result

Method: Theta method Recommended

Week 1 (Oct 4, 2026): 1.9%, 80% band 1.76% – 2.06%
Week 8 (Nov 22, 2026): 3.26%, 80% band 2.52% – 4.2%

Forecast shown

Basis and limits

Load data
Pass
835 weeks (Oct 3, 2010 – Sep 27, 2026), 0 missing
Data checks
Pass
No findings.
Rolling-origin backtest
Pass

52 origins over 3.1 seasons

Chose AutoTheta: MAE is 0.58× the seasonal-naive baseline (threshold < 0.95); 80% band covered 81%
Decision
Forecast allowed

forecast_allowed: true reason: beats_seasonal_naive_in_rolling_origin_backtest reason_codes: []

US national ILI outpatient visits

Decision Forecast allowed

② Data overview

Time span: Oct 3, 2010 – Sep 27, 2026
Frequency: One value per week
Length: 835 weeks (≈ 16.1 season(s))
Latest value: 47,389 (Sep 27, 2026)
Last 52 weeks: 20,242 – 215,792
Missing: 0 weeks (0%), longest gap 0 weeks
Seasonality: Clear: strength 0.81 (cut-off 0.35)

Download this series JSON Parquet

History
Forecast
80% band / 95% band

The chart shows the last 104 weeks; the blank area on the right is the future to forecast.

Data check-up

NoteNon-stationary after removing seasonality (ADF p=0.609, KPSS p=0.010, level drift +1.72): this is hard to judge; seasonal epidemics alone do not trigger this, a trend or random walk does.

③ Choose a forecasting method

Each method was replayed at 52 past points without seeing what came next, forecasting 8 weeks ahead each time (rolling backtest). Shorter bar = smaller average error.

Choose a forecasting method Theta methodRecommended Removes the season, smooths the trend, then adds the season back; parameters chosen automatically. Mean error 13,264 · 33% below baseline Lowest backtest error: 33% below repeating last year. MSTL decomposition Splits the series into seasonal and trend parts and extends each. Mean error 14,307 · 28% below baseline Error 8% higher than the recommended method. Naive Carries the latest week forward. Mean error 24,005 · 20% above baseline Did not beat the baseline (1.2× its error); not used on its own. Seasonal naive (baseline) Repeats the same week last year; every other method has to beat it. Mean error 19,930 Baseline, for comparison only.

Unavailable methods (1)

Seasonal ETS (exponential smoothing)Unavailable Season length is 52 weeks: statsforecast cannot fit a seasonal ETS that long (it collapses to Naive), so it is not a candidate.

④ Forecast result

Method: Theta method Recommended

Week 1 (Oct 4, 2026): 49,495, 80% band 44,753 – 54,740
Week 8 (Nov 22, 2026): 86,330, 80% band 62,027 – 120,153

Forecast shown · Recommended: The forecast passed the backtest, but the series is hard to judge (non-stationary / unstable); treat horizons beyond the first few weeks with care.

Basis and limits

Load data
Pass
835 weeks (Oct 3, 2010 – Sep 27, 2026), 0 missing
Data checks
Warning
- trend_nonstationary warn
Rolling-origin backtest
Pass

52 origins over 3.1 seasons

Chose AutoTheta: MAE is 0.67× the seasonal-naive baseline (threshold < 1); 80% band covered 81%
Decision
Forecast allowed

forecast_allowed: true reason: beats_seasonal_naive_in_rolling_origin_backtest reason_codes: [trend_nonstationary]

England: influenza test positivity (%)

Decision Forecast limited

② Data overview

Time span: Jul 3, 2017 – Sep 28, 2026
Frequency: One value per week
Length: 483 weeks (≈ 9.3 season(s))
Latest value: 5.64% (Sep 28, 2026)
Last 52 weeks: 0.34% – 21.4%
Missing: 0 weeks (0%), longest gap 0 weeks
Seasonality: Clear: strength 0.62 (cut-off 0.35)

Download this series JSON Parquet

History
Forecast
80% band / 95% band

The chart shows the last 104 weeks; the blank area on the right is the future to forecast.

Data check-up

Limits the forecast40 outlying periods (8.3%, max deviation 2.70 model units) distort fitting and interval calibration. Consider masking them.
NoteA run of 9 exact zeros inside an otherwise active series (possible implicit zero-fill of missing data).

③ Choose a forecasting method

Each method was replayed at 52 past points without seeing what came next, forecasting 8 weeks ahead each time (rolling backtest). Shorter bar = smaller average error.

The check-up limits this series: whichever method, at most 4 weeks ahead and bands widened ×1.25.

Choose a forecasting method MSTL decompositionRecommended Splits the series into seasonal and trend parts and extends each. Mean error 2.3 pp · 33% below baseline Lowest backtest error: 33% below repeating last year. Naive Carries the latest week forward. Mean error 3 pp · 12% below baseline Error 30% higher than the recommended method. Theta method Removes the season, smooths the trend, then adds the season back; parameters chosen automatically. Mean error 3.1 pp · 10% below baseline Error 35% higher than the recommended method. Seasonal naive (baseline) Repeats the same week last year; every other method has to beat it. Mean error 3.4 pp Baseline, for comparison only.

Unavailable methods (1)

Seasonal ETS (exponential smoothing)Unavailable Season length is 52 weeks: statsforecast cannot fit a seasonal ETS that long (it collapses to Naive), so it is not a candidate.

④ Forecast result

Method: MSTL decomposition Recommended

Week 1 (Oct 5, 2026): 5.99%, 80% band 3.26% – 10.7%
Week 4 (Oct 26, 2026): 7.44%, 80% band 2.8% – 18.3%

Forecast shown with limits: horizon ≤ 4 weeks, intervals ×1.25 · Recommended: Mask known disruption periods (e.g. COVID-19) via `mask_periods` so they do not affect fitting or interval calibration.

Basis and limits

Load data
Pass
483 weeks (Jul 3, 2017 – Sep 28, 2026), 0 missing
Data checks
Limit
- outlier_contamination limit
- zero_runs warn
Rolling-origin backtest
Pass

52 origins over 3.1 seasons

Chose MSTL: MAE is 0.67× the seasonal-naive baseline (threshold < 0.95); 80% band covered 81%
Decision
Forecast limited

forecast_allowed: true reason: beats_seasonal_naive_in_rolling_origin_backtest reason_codes: [outlier_contamination, zero_runs]

EU/EEA influenza positivity, sentinel (%)

Decision Forecast refused

② Data overview

Time span: Jun 20, 2022 – Oct 5, 2026
Frequency: One value per week
Length: 225 weeks (≈ 4.3 season(s))
Latest value: 4% (Oct 5, 2026)
Last 52 weeks: 0% – 47%
Missing: 0 weeks (0%), longest gap 0 weeks
Seasonality: Clear: strength 0.88 (cut-off 0.35)

Download this series JSON Parquet

History

The chart shows the last 104 weeks.

Data check-up

Note8 outlying periods (3.6%, max deviation 4.10 model units) distort fitting and interval calibration. Consider masking them.

This series is not suitable for forecasting

The backtest covers only 2.3 seasons (at least 3 needed), too few to tell whether it beats the baseline.

Basis and limits

Load data
Pass
225 weeks (Jun 20, 2022 – Oct 5, 2026), 0 missing
Data checks
Warning
- outlier_contamination warn
Rolling-origin backtest
Block
38 origins over 2.3 seasons

No candidate model could be chosen.
- insufficient_backtest block Backtest too small to judge skill: 38 origins spanning 2.3 seasons (need 10 origins).
Decision
Forecast refused

forecast_allowed: false reason: backtest_lt_3_seasons reason_codes: [insufficient_backtest, outlier_contamination]

Malta: influenza detections, sentinel

Decision Forecast refused

② Data overview

Time span: Oct 24, 2022 – Oct 9, 2023
Frequency: One value per week
Length: 51 weeks (≈ 1 season(s))
Latest value: 2 (Oct 9, 2023)
Last 52 weeks: 1 – 3
Missing: 41 weeks (80%), longest gap 39 weeks
Seasonality: History too short to estimate

Download this series JSON Parquet

History

The chart shows the last 51 weeks.

Data check-up

Blocks the forecastHistory is too short: 51 periods = 1.0 seasons (at least 3 needed).
Blocks the forecastToo many missing values: 80.4% missing, longest internal gap 39 periods (limits: 25% / 2 periods).

This series is not suitable for forecasting

Not enough data: 51 weeks (≈ 1 season(s)), at least 3 seasons needed.
Too many gaps: 80% missing, longest gap 39 weeks (limit 2).

Basis and limits

Load data
Pass
51 weeks (Oct 24, 2022 – Oct 9, 2023), 41 missing
Data checks
Block
- too_short block
- too_many_gaps block
Rolling-origin backtest
Skipped
Not run: a data check blocked, so there is no backtest.
Decision
Forecast refused

forecast_allowed: false reason: history_lt_3_seasons reason_codes: [too_short, too_many_gaps]

Shared by all series: methods tried in the offline benchmark, not production candidates

Seasonal window average (3 seasons)Unavailable Not a production candidate. Offline benchmark (BENCHMARK.md, 11 scenarios pooled): error 0.94× seasonal naive vs 0.74× for the current selection rule; COVID-like synthetic scenario 2.38× vs 0.91×. US national weighted ILI (wILI, %): 0.58× (current 0.59×)US national ILI outpatient visits: 0.54× (current 0.66×) This series is not in the benchmark; only the pooled figures above apply. MSTL with AutoARIMA trendUnavailable Not a production candidate. Offline benchmark (BENCHMARK.md, 11 scenarios pooled): error 0.76× seasonal naive vs 0.74× for the current selection rule; COVID-like synthetic scenario 0.92× vs 0.91×. US national weighted ILI (wILI, %): 0.68× (current 0.59×)US national ILI outpatient visits: 0.72× (current 0.66×) This series is not in the benchmark; only the pooled figures above apply. Ensemble: MSTL + ThetaUnavailable Not a production candidate. Offline benchmark (BENCHMARK.md, 11 scenarios pooled): error 0.77× seasonal naive vs 0.74× for the current selection rule; COVID-like synthetic scenario 1.13× vs 0.91×. US national weighted ILI (wILI, %): 0.61× (current 0.59×)US national ILI outpatient visits: 0.7× (current 0.66×) This series is not in the benchmark; only the pooled figures above apply. Ensemble: MSTL + Theta + seasonal window averageUnavailable Not a production candidate. Offline benchmark (BENCHMARK.md, 11 scenarios pooled): error 0.78× seasonal naive vs 0.74× for the current selection rule; COVID-like synthetic scenario 1.19× vs 0.91×. US national weighted ILI (wILI, %): 0.56× (current 0.59×)US national ILI outpatient visits: 0.63× (current 0.66×) This series is not in the benchmark; only the pooled figures above apply. Ensemble: MSTL + Theta + naiveUnavailable Not a production candidate. Offline benchmark (BENCHMARK.md, 11 scenarios pooled): error 0.81× seasonal naive vs 0.74× for the current selection rule; COVID-like synthetic scenario 1.61× vs 0.91×. US national weighted ILI (wILI, %): 0.67× (current 0.59×)US national ILI outpatient visits: 0.78× (current 0.66×) This series is not in the benchmark; only the pooled figures above apply.

Reproduce

uv run python scripts/build_showcase_data.py
jq '.backtest.models | map_values(.mae)' \
  site/public/data/series/<series_id>.json
jq '.gate | {decision, reason_codes}' \
  site/public/data/series/<series_id>.json

Raw JSON (manifest)

Results come from static build-time data, not a live computation; a refusal is not automatically right. Not for clinical or public-health decisions.

← Back to the home page

Run a forecast

① Choose the series to forecast

② Data overview

Data check-up

③ Choose a forecasting method

④ Forecast result

Load data

Data checks

Rolling-origin backtest

Decision

② Data overview

Data check-up

③ Choose a forecasting method

④ Forecast result

Load data

Data checks

Rolling-origin backtest

Decision

② Data overview

Data check-up

③ Choose a forecasting method

④ Forecast result

Load data

Data checks

Rolling-origin backtest

Decision

② Data overview

Data check-up

This series is not suitable for forecasting

Load data

Data checks

Rolling-origin backtest

Decision

② Data overview

Data check-up

This series is not suitable for forecasting

Load data

Data checks

Rolling-origin backtest

Decision