Approach of the yield forecast

(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)

The forecasted crop yield is calculated with a combination of methods and software tools. It is assumed to be a function of the trend of observed yields, possibly corrected with a function of one or more indicators from elsewhere out of the MCYFS and closed with a residual error:

Forecasted yield = f(t) + f(i) + residual error

Where:

• f(t) = function of the trend of observed yields over the last 10 years
• f(i) = optional function of one or more indicators that originate from other parts of the MCYFS.

The approach of using different combinations of indicators results in a large number of possible forecasts with only a few being statistically significant and even fewer having a logical causal relation between indicator and observed yield.

Priorities in selecting the forecast model

Because there can be several significant forecast models with a logical causal relation between indicator and observed yield, priorities are followed in selecting the 'best' model.

1. As a first step, yields are predicted by the CGMS statistical sub system which uses agricultural yield statistics and simulated crop indicators. This system consists of a linear regression model combining the mean yield, a linear time trend and a linear regression function to explain the residual variation (Vossen, 1990b; 1992). The linear time trend represents the influence of long-term economic and technological dynamics such as increased fertiliser application, improved crop management methods, new high yielding varieties, etc. on yields. The residual variation is modelled as a function of crop growth simulation results (potential or water limited dry weight of the simulated biomass or storage organs). As said before these simulated crop indicators account for the inter-annual yield variation that results from weather variability.
2. When for a certain combination of country and crop the accuracy of the predicted yield is deemed not to be sufficient, the MARS analysts at JRC start to redefine trend periods and functions. This part must not be underestimated as the trend reference can have a serious effect on the final forecasting model selected.
3. In some occasions the analysts build their own prediction models for certain combinations of crops and countries. These models use other crop indicators like simulated leaf area index and CNDVI.
4. To deal with the uncertainty between the moment the forecast is issued and the moment the crop is harvested, agro-meteorological scenario's can be produced and analysed. The scenarios are based on agro-meteorological similar years as detected by cluster and factorial analyses (PCA techniques). Once agro-climatic years are detected which are similar to the current year, the resulting range of final yield performances of these years can be attributed to the current season. The extremes maximum and minimum yields obtained in the clusters of similar years are always characterised through the factorial analysis. If a trend exists the range of final yields will be corrected for this trend before relating the yields to the current year. This technique helps to understand how the yield prediction could still change before harvesting.
5. As a last attempt to determine the yield forecast, and in the optic to be able to release a forecast at EU level by aggregating National forecasts, the calculation of the trimmed average is performed. I.e. yield statistics of the last five years are taken and ranked. The trimmed average is the average of the three centre years. This trimmed average will be used as yield forecast when all other methods (prediction models such as CGMS level 3, trend analysis, scenario analysis) do not lead to satisfactory results or when there is a lack of data, a gap in the time series.

The final predicted yields for each country, as established by the MARS analysts, are published in the MARS bulletin available on Bulletins-Publications.

Accuracy

According to previous analysis (Genovese 1998) the desired level of accuracy is reached when the error is lower than 3% at national level. The level remains appreciable if it is lower than 6%. If the error is greater than 6% the yield forecast can be considered unacceptable with one exception: if the error is still lower than that of any other source available. In that case the MCYFS still produces an improvement which could be economically relevant. For winter crops the yield forecast should be accurate before June, for summer crops before August.

A global evaluation of the system performances was made in 1998 using data until 1997 (Genovese, 1998). The RMSE error has been used as error indicator by comparing the MCYFS forecasts with ex-post observed results. The average yield forecast error ranges from about 3% to 5% at EU level for the main crops (8.3% for durum wheat). In absolute terms the average error for wheat ranges between about 2 and 4 quintals ha -1 both for European and national levels (Genovese, 2001).

In general what has been shown is that the system gives higher errors at the beginning of the season and lower at the end according to a cumulative effect of the climate impact on the crop behaviour. The error analysis is repeated each year in order to carefully watch the performance of the MCYFS, individuate sources of errors and take decision on which part of the MCYFS has to be improved. The MARS Unit launched a study called QUAMP to calculate the overall performance of the system including the period 1998-2001.