How good is each of the brain's forecasts against reality?

The brain publishes a forecast only when its method beat a naive baseline out of sample, such as the same month last year, last year's rate or a district's own average, under a protocol committed before any scoring, with 80% intervals that held. Most methods tried so far failed that test, so most of what the brain has built stays a backtest: its scores are shown, its forecasts are not. A forecast published live is written to an append-only, hash-chained log before its outcome exists and scored here when the outcome is published, so its record cannot be edited afterwards.

Across the backtests below, 8 of 28 candidate model rows passed their gate. Statuses as of 10 Oct 2026:

2Live Logged before the outcome existed, published, and scored here when the outcome arrives.
0Published, not logged A publish verdict with a target still open, but the forecasts were not logged before their outcomes, so they cannot be scored live.
2Backtest only A model passed its gate, but every forecast target has passed: backtest scores only.
2Withheld No model beat the best baseline under the pre-registered gate: backtest scores only.
1Not validated The method failed its pre-registered gate: backtest scores only, nothing published.
1Not published (logging) Pre-registered with its live log open, but no forecast origin exists yet: nothing is logged, scored or published.

District neonatal mortality rate (NMR), one to three months ahead

Live

What will each district's facility-reported neonatal mortality rate be in the months DGHS has not printed yet?

104 district nowcasts logged before DGHS printed their month (Sep 2025 and Oct 2025) are published.

Last checked: 1 Oct 2026, 02:58 UTC: the latest DGHS fetch recorded in the NMR live log.
Next due: 24 Oct 2026, 21:02 UTC: adoption re-check of Nov 2025 and Dec 2025 under the 30-day stability rule: no month is adopted before 30 days after its first fetch printing its current values in every division (Nov 2025 from 24 Oct 2026, 21:02:08 UTC; Dec 2025 from 24 Oct 2026, 21:02:47 UTC).

Pre-registration: Filed 24 Sep 2026, 17:56 UTC. Committed before any model was fitted or scored; an erratum was recorded. Protocol on the product page
Target: NMR as printed by the DGHS RMNCH Scorecard for a district and month (0.0 treated as missing); nowcasts of months the source had not printed when the forecast was logged.
Gate: Beats the best of three baselines (seasonal naive, last value, district mean) on MAE in both the development and the test window, 80% interval coverage within 0.65 to 0.95, at least 12 test origins per horizon.
Baselines: Seasonal naive (same month last year); Last observed value; District mean to date.

Backtest skill against the baselines

Test window, targets Sep 2024 to Aug 2025, horizons 1 to 3 pooled. Relative MAE is the MAE over the MAE of the best baseline (district mean to date); below 1 beats it.

Model	Verdict	Relative MAE	80% coverage	Points
Pooled shrinkage across districtsPasses	Passes	0.840	0.797	1,488
Exponential smoothing, ETS(A,N,N)Passes	Passes	0.923	0.789	1,488
District mean to date (baseline)Best baseline	Best baseline	1.000	0.817	1,488
Last observed value (baseline)Baseline	Baseline	1.148	0.778	1,488
Seasonal naive (same month last year) (baseline)Baseline	Baseline	1.163	0.767	1,488

Live record

Forecasts logged: 156
Logged before their outcome: 104
Published: 104 (Sep 2025; Oct 2025)
Baselines logged beside them: 104, method value reproduced for 104
Outcomes recorded: 0
Outcomes scored: 0

No outcomes scored yet. Each logged target is scored here once the source prints it and the refresh adopts it.

Log: 213 entries, hash chain verified on this read; last source check 1 Oct 2026, 02:58 UTC. Registry: 109 entries, 3 anchors matching.

Open Health early warning

District Full Vaccination coverage, one to three months ahead

Not validated

Which districts' Full Vaccination coverage will fall below 80%?

The pre-registered gate failed: no method beat the best baseline in both windows.

Pre-registration: Filed 24 Sep 2026, 17:56 UTC. Committed before any model was fitted or scored; an erratum was recorded. Protocol on the product page
Target: Full Vaccination coverage (% as printed, can exceed 100) for a district and month.
Gate: Beats the best of three baselines (seasonal naive, last value, district mean) on MAE in both the development and the test window, 80% interval coverage within 0.65 to 0.95, at least 12 test origins per horizon.
Baselines: Seasonal naive (same month last year); Last observed value; District mean to date.

Backtest skill against the baselines

Test window, targets Sep 2024 to Aug 2025, horizons 1 to 3 pooled. Relative MAE is the MAE over the MAE of the best baseline (district mean to date); below 1 beats it.

Model	Verdict	Relative MAE	80% coverage	Points
Pooled shrinkage across districtsFails	Fails	1.025	0.653	1,920
Exponential smoothing, ETS(A,N,N)Fails	Fails	1.060	0.667	1,920
District mean to date (baseline)Best baseline	Best baseline	1.000	0.667	1,920
Seasonal naive (same month last year) (baseline)Baseline	Baseline	1.088	0.646	1,920
Last observed value (baseline)Baseline	Baseline	1.308	0.635	1,920

Live record

No live log: this problem's forecasts were never logged before their outcomes, so none can be scored live.

Open Health early warning

Monthly actual spending by ministry and division (P2)

Backtest only

What will each ministry's operating and development spending be in the next one to three months?

26 model verdicts pass the gate, but every forecast target has passed (the last ended 31 Jan 2026).

Protocol recorded after the fact; not a pre-registration: The protocol is the text of the project plan dated 23 Sep 2026, added after scoring; no pre-registration was committed before scoring. Protocol on the product page
Target: Monthly actuals in Tk crore per printed operating or development line of the Finance Division's monthly fiscal report.
Gate: MAE below seasonal naive out of sample (horizons 1 to 3 pooled), 80% interval coverage within 0.65 to 0.95, at least 12 origins.
Baselines: Seasonal naive (same month last year). The backtest read the same version of the source data that is served here.

Backtest skill against the baselines

Rolling-origin held-out targets, horizons 1 to 3. Relative MAE is the MAE over the MAE of seasonal naive (same month last year); below 1 beats it. Decided per series: 25 of 122 series have a model that passes (26 model verdicts); the pooled rows below are context.

Group	Model	Verdict	Relative MAE	80% coverage	Points
Development	developmentExponential smoothing (seasonal)Fails	Fails	1.030	0.773	1,889
Development	developmentLast month carried forwardFails	Fails	1.373	0.769	1,891
Development	developmentSeasonal naive (same month last year) (baseline)Baseline	Baseline	1.000	0.666	1,891
Development	developmentSeasonal naive with year-to-date growthFails	Fails	1.471	0.730	1,708
Operating	operatingExponential smoothing (seasonal)Passes	Passes	0.935	0.745	1,890
Operating	operatingLast month carried forwardFails	Fails	1.176	0.714	1,915
Operating	operatingSeasonal naive (same month last year) (baseline)Baseline	Baseline	1.000	0.714	1,915
Operating	operatingSeasonal naive with year-to-date growthFails	Fails	2.820	0.717	1,879

Live record

No live log: this problem's forecasts were never logged before their outcomes, so none can be scored live.

Open Predict: P2

e-GP contract awards per month (P5a)

Backtest only

How many contracts will be awarded on e-GP next month, nationally, by ministry and by district?

74 model verdicts pass the gate, but no forecast was logged before its outcome, so they are backtest only (not logged before outcome); logging starts from the next forecast origin.

Protocol recorded after the fact; not a pre-registration: The protocol is the text of the project plan dated 23 Sep 2026, added after scoring; no pre-registration was committed before scoring. Protocol on the product page
Target: Count of deduplicated e-GP awards dated in the month, per series as printed.
Gate: MAE below seasonal naive out of sample (horizons 1 to 3 pooled), 80% interval coverage within 0.65 to 0.95, at least 12 origins.
Baselines: Seasonal naive (same month last year). The backtest read the same version of the source data that is served here.

Backtest skill against the baselines

Rolling-origin held-out targets, horizons 1 to 3. Relative MAE is the MAE over the MAE of seasonal naive (same month last year); below 1 beats it. Decided per series: 58 of 140 series have a model that passes (74 model verdicts); the pooled rows below are context.

Group	Model	Verdict	Relative MAE	80% coverage	Points
District	districtExponential smoothing (seasonal)Passes	Passes	0.887	0.675	2,277
District	districtLast month carried forwardFails	Fails	1.314	0.623	2,277
District	districtSeasonal naive (same month last year) (baseline)Baseline	Baseline	1.000	0.691	2,277
District	districtSeasonal naive with year-to-date growthFails	Fails	1.106	0.660	2,112
Ministry	ministryExponential smoothing (seasonal)Passes	Passes	0.914	0.678	2,193
Ministry	ministryLast month carried forwardFails	Fails	1.123	0.673	2,304
Ministry	ministrySeasonal naive (same month last year) (baseline)Baseline	Baseline	1.000	0.642	2,304
Ministry	ministrySeasonal naive with year-to-date growthFails	Fails	1.510	0.605	1,710
National	nationalExponential smoothing (seasonal)Fails	Fails	0.991	0.576	33
National	nationalLast month carried forwardFails	Fails	1.569	0.515	33
National	nationalSeasonal naive (same month last year) (baseline)Baseline	Baseline	1.000	0.545	33
National	nationalSeasonal naive with year-to-date growthFails	Fails	0.928	0.515	33

Live record

Forecasts logged: 0
Logged before their outcome: 0
Published: 0
Baselines logged beside them: 0
Outcomes recorded: 0
Outcomes scored: 0

No outcomes scored yet. Each logged target is scored here once the source prints it and the refresh adopts it.

Log: 0 entries, hash chain verified on this read. Registry: 0 entries, 0 anchors matching. No P5a forecast has been logged yet: logging starts from the next forecast origin whose target months have not begun.

Open Predict: P5a

ADP execution by ministry at year-end, from the IMED monthly edition (P3b)

Withheld

Which ministries will under-spend their ADP allocation by the end of the fiscal year?

No candidate model passed the pre-registered gate against the best baseline.

Pre-registration: Filed 24 Sep 2026, 12:22 UTC. Committed before scoring. Protocol on the product page
Target: June expenditure over the June allocation per ministry or division, as IMED prints it.
Gate: MAE below the best baseline (last year's rate; even spread and seasonal naive also shown) on the development and the final fiscal year, 80% interval coverage within 0.65 to 0.95, at least 12 origins, and the served origin month's own check.
Baselines: Last year's rate; Even spread through the year; Seasonal naive (same month last year).

Backtest skill against the baselines

Held-out fiscal years. Relative MAE is the MAE over the MAE of last year's rate; below 1 beats it.

Group	Model	Verdict	Relative MAE	80% coverage	Points
Ministry	ministryEven spread through the year (baseline)Baseline	Baseline	3.094	0.721	542
Ministry	ministrySeasonal naive (same month last year) (baseline)Baseline	Baseline	66.807	0.761	542
Ministry	ministryLast year's rate (baseline)Baseline	Baseline	1.000	0.546	542
Ministry	ministryRemainder of allocation, spreadFails	Fails	1.132	0.615	542
Ministry	ministryPooled quantile of past yearsFails	Fails	0.978	0.685	542
Ministry	ministryShrinkage toward the pooled profileFails	Fails	1.017	0.774	542

Live record

No live log: this problem's forecasts were never logged before their outcomes, so none can be scored live.

Open Predict: P3b board

ADP execution at year-end, from the April edition on (P3b-late)

Withheld

From the April edition on, which ministries will under-spend their ADP allocation by year-end?

No candidate model passed the pre-registered gate on the unseen Finance Division years (FY2014-15, FY2015-16).

Pre-registration: Filed 24 Sep 2026, 13:44 UTC. Committed before any Finance Division back year was scored. Protocol on the product page
Target: June expenditure over the June allocation per ministry, forecast from months 9, 10 and 11.
Gate: MAE below last year's rate on the unseen Finance Division years (FY2014-15 and FY2015-16, each year too), 80% coverage within 0.65 to 0.95, at least 12 origins and 12 interval targets.
Baselines: Last year's rate; Even spread through the year; Seasonal naive (same month last year).

Backtest skill against the baselines

Confirmatory Finance Division years (FY2014-15, FY2015-16), origins at months 9 to 11. Relative MAE is the MAE over the MAE of last year's rate; below 1 beats it.

Model	Verdict	Relative MAE	80% coverage	Points
Even spread through the year (baseline)Baseline	Baseline	4.648	0.964	28
Seasonal naive (same month last year) (baseline)Baseline	Baseline	5.789	0.833	28
Last year's rate (baseline)Baseline	Baseline	1.000	0.333	28
Remainder of allocation, spreadFails	Fails	1.638	0.889	28
Shrinkage toward the pooled profileFails	Fails	1.240	1.000	28
Pooled median of past yearsFails	Fails	1.365	1.000	28

Live record

No live log: this problem's forecasts were never logged before their outcomes, so none can be scored live.

Open Predict: P3b-late board

Quarterly real GDP growth nowcast from Bangladesh Bank's monthly indicators

Live

How fast is Bangladesh's economy growing in the quarter BBS has not printed yet?

1 quarterly nowcast logged before BBS printed the quarter is published: FY2025-26 Q4, Apr to Jun 2026.

Last checked: 1 Oct 2026, 02:58 UTC: the latest BBS GDP release listing check recorded in the nowcast live log.
Next due: Not scheduled: the BBS GDP listing is checked by hand; FY2025-26 Q4, Apr to Jun 2026 unprinted at the last check.

Pre-registration: Filed 25 Sep 2026, 00:40 UTC. Committed before any model was fitted or scored; amendment 1 (25 Sep 2026, 00:44 UTC, also before any score) reads every predictor as of the origin. Protocol on the product page
Target: Y-o-y growth of quarterly real GDP at constant 2015-16 prices as BBS prints it: the latest print in the backtest, the first print (the release named for the quarter) for live nowcasts.
Gate: Pooled MAE below the best of three baselines (last value, historical mean, AR(1)) on the common scored origins within 2019-20 Q1 to 2025-26 Q3 and in the final window from 2023-24 Q1, 80% interval coverage within 0.65 to 0.95, at least 16 common origins.
Baselines: Last observed value; Historical mean of GDP growth; AR(1) on GDP growth.

Backtest skill against the baselines

20 quarters scored by every method, Apr to Jun 2021 to Jan to Mar 2026, against the latest BBS print; each nowcast from the editions out by its origin. Relative MAE is the MAE over the MAE of the best baseline (historical mean of GDP growth); below 1 beats it.

Model	Verdict	Relative MAE	80% coverage	Points
One-factor model (six indicators)Passes	Passes	0.552	0.900	20
Equal-weight composite (five indicators)Passes	Passes	0.745	0.750	20
Bridge regression (exports, remittances, private credit)Passes	Passes	0.963	0.900	20
Historical mean of GDP growth (baseline)Best baseline	Best baseline	1.000	0.950	20
AR(1) on GDP growth (baseline)Baseline	Baseline	1.013	0.950	20
Last observed value (baseline)Baseline	Baseline	1.158	0.900	20

Live record

Forecasts logged: 1
Logged before their outcome: 1
Published: 1 (FY2025-26 Q4, Apr to Jun 2026)
Baselines logged beside them: 1
Outcomes recorded: 0
Outcomes scored: 0

FY2025-26 Q4, Apr to Jun 2026: 3.04% (percent, y-o-y real GDP growth; 80% interval -0.57% to 6.66%), logged 25 Sep 2026, 01:35 UTC; the best baseline logged beside it: 5.73%. Awaiting the source's first print.

Chain verified: 9 entries in the live log, 109 in the registry log.

Logged: 25 Sep 2026, 01:35 UTC
Input edition: Jun 2026 edition, released 27 Jul 2026
Pre-registered: 25 Sep 2026, 00:40 UTC

Logged: 25 Sep 2026, 01:35 UTC
Input edition: Jun 2026 edition, released 27 Jul 2026
Pre-registered: 25 Sep 2026, 00:40 UTC

No outcomes scored yet. Each logged target is scored here once the source prints it and the refresh adopts it.

Log: 9 entries, hash chain verified on this read; last source check 1 Oct 2026, 02:58 UTC. Registry: 109 entries, 2 anchors matching.

Open Economy nowcast

DGDA medicine price per unit by generic, one to three months ahead (P6)

Not published (logging)

What will the DGDA maximum retail price per unit of each generic medicine be in the next three months?

No October 2026 observations yet: nothing logged (no origin: the source holds 3 fetch runs (27 Sep 2026 at 21:58 UTC, 27 Sep 2026 at 22:00 UTC, 29 Sep 2026 at 01:25 UTC), observing 1 month (Sep 2026); an origin needs a generic observed in at least 2 months, the last of them ended).

Last checked: 1 Oct 2026, 03:07 UTC: the latest P6 rebuild of the series from DGDA's price pages, or forecast logged, whichever is later.
Next due: 1 Nov 2026, 00:00 UTC: next origin Oct 2026, if DGDA is fetched in October 2026: loggable from 1 Nov 2026, horizon 2 before 1 Dec 2026.

Pre-registration: Filed 1 Oct 2026, 03:02 UTC. Committed before the series builder existed, before any model was fitted and before anything was scored. Protocol on the product page
Target: Monthly median unit price (Tk per unit) per generic over a fixed basket of DGDA products seen in at least 2 fetch runs, from DGDA's price-detail pages as fetched.
Gate: MAE below seasonal naive (pooled across generics, horizons 1 to 3 pooled), 80% interval coverage within 0.65 to 0.95, at least 12 distinct origin months; a seasonal-naive MAE of 0 withholds.
Baselines: Last observed value; Seasonal naive (same month last year).

Backtest skill against the baselines

No backtest: the pre-registered gate needs 12 origin months and no origin exists yet.

Live record

Forecasts logged: 0
Logged before their outcome: 0
Published: 0
Baselines logged beside them: 0
Outcomes recorded: 0
Outcomes scored: 0

No outcomes scored yet. Each logged target is scored here once the source prints it and the refresh adopts it.

Log: 0 entries, hash chain verified on this read; last source check 1 Oct 2026, 03:07 UTC. Registry: 109 entries, 0 anchors matching.