Cereal yield · World · 2 years ahead
2026: the model assigns an 80 percent probability to 3,927 to 4,113 kg per hectare
Point estimate 4,026 kg per hectare from Ridge regression on lags and drivers, against 3,955 kg per hectare observed in 2024. ridge beats last_value on out-of-sample MAE by 29.3 percent over 15 cutoffs (minimum 8, margin 5 percent).
Fan chart
- Quantiles
- p5 3,828 · p10 3,927 · p25 3,975 · p50 4,006 · p75 4,045 · p90 4,113 · p95 4,176
- Source
- empirical quantiles of 12 in-window 2-step residuals, in log space; no normality assumed
Immutable fields
- Key
- cereal-yield:WLD:2
- Id
- c8663db1-8952-4fe8-8d4b-cee53cee8d45
- Created
- 2026-09-07 22:16 UTC
- Data cutoff
- 2026-09-07 22:16 UTC
- Origin
- 2024 = 3,955 (obs 2945)
- Target period
- 2026 (731 days)
- Point
- 4,026 kg per hectare
- Interval
- 3,927 to 4,113 at 80 percent
- Model
- ridge models_v1.targets_v1
- Baseline
- last_value
- Snapshot
- fd49ea004ce91b4ec69de2495494ec6d626e0ed332a88ae95263a08ec68e68b4
- Data mode
- ingested_at
- Status
- active
Statistical association, not causal
Ridge coefficient times the standardised input at the origin, in log kg per hectare. These are associations learned in the training window, labelled as such; nothing here is a causal effect.
| Feature | Input | Coef | Contribution |
|---|---|---|---|
| cereal-production[t-1] | 3,126,272,362 | -0.0023 | -0.0046 |
| y[t-0] | 8.28 | -0.0033 | -0.0043 |
| y[t-1] | 8.35 | -0.0020 | -0.0031 |
| world-population[t-1] | 8,062,923,417 | -0.0013 | -0.0024 |
| agricultural-land-area[t-1] | 4,642,071 | -0.0010 | +0.0004 |
Not yet realised
The observation for 2026 has not been ingested. The weekly forecast.evaluate job attaches it when it arrives; the record above does not change.
What the forecast takes for granted
- No structural break in cereal-yield between the data cutoff and the target period.
- The source keeps publishing the series on the same definition; a rebasing or redefinition invalidates the comparison.
- The lagged statistical associations estimated in the training window hold over the horizon.
- Driver series used, lagged: world-population, agricultural-land-area, cereal-production. Their own future values are not forecast; only values known at the cutoff enter.
What would show this forecast wrong
- If the first published value of cereal-yield (WLD) for 2026 is below 3,927 kg per hectare or above 4,113 kg per hectare, the 80 percent interval is falsified.
- The model assigns a 10 percent probability to a value below 3,927 kg per hectare and a 10 percent probability to a value above 4,113 kg per hectare; interval misses should occur about one time in five over many forecasts, and a run of misses well above that rate falsifies the calibration.
- The point forecast is above the last observed value (3,955 kg per hectare); a realised value at or below that level falsifies the direction call.
Scorecard for this target and horizon
| Model | Status | Cutoffs | MAE | RMSE | Direction | 80% coverage | Skill vs last value |
|---|---|---|---|---|---|---|---|
| Ridge regression on lags and drivers models_v1.targets_v1 | champion | 15 | 72.74 | 111.3 | 93% | 87% | 29.3% |
| Drift models_v1.targets_v1 | challenger | 15 | 99.22 | 131 | 93% | 80% | 3.6% |
| Holt (exponential smoothing with additive trend) models_v1.targets_v1 | challenger | 15 | 81.85 | 112.9 | 93% | 80% | 20.5% |
| Last value models_v1.targets_v1 | baseline | 15 | 102.9 | 127 | 0% | 87% | 0.0% |
| Linear trend models_v1.targets_v1 | challenger | 15 | 90.98 | 121.9 | 87% | 67% | 11.6% |
Ridge regression on lags and drivers
Ridge regression on lags and drivers
Purpose. Point and 80 percent interval forecasts of cereal-yield (WLD) 1, 2, 5 years ahead. Owner: Forecast Lab (HV 3.0).
Method
A direct multi-horizon ridge regression: the change from the origin to the target period is regressed on lagged target values (lags [1,2]) and lagged driver series (lags [1]), features standardised, penalty lambda 1. Drivers: world-population, agricultural-land-area, cereal-production. Every driver enters lagged, so no feature uses a period after the origin. The series is modelled in natural logarithms and transformed back, so intervals are asymmetric in level terms.
Uncertainty
Intervals are empirical: the model is refitted at earlier origins inside the training window, the realised errors at the same horizon are collected, and the interval is the 10th to 90th percentile of those errors around the point forecast. No normality is assumed.
Training and evaluation
Training window: 64 periods, 1961-01-01 to 2024-01-01. Evaluation: rolling-origin backtest over 15 cutoffs using pseudo real-time data (the earliest ingested value per period; no vintage archive exists for this series, so revisions between the cutoff and today are not reproduced).
Known limitations
- Structural breaks after the cutoff are not modelled.
- Driver contributions are statistical associations, not causal effects.
- Promotion to champion requires beating the last-value baseline on out-of-sample MAE by at least 5 percent over the minimum number of cutoffs; otherwise the baseline stays champion.
- Retraining: every scheduled forecast.run refits on the data then available; backtests re-run from /admin/forecasts.
Supersedes
This is the first record for its key.