which is why there are no pixels represented from the country. The locations with related NDVI
pixels include a dense cluster south of Lake Cahora Bassa, which the FEWS NET map describes
as a minor deficit zone, regions in the south of the country in Chokwe, and regions in the north
in Tanzania near the Rift Valley lakes. These maps of active NDVI pixels, once included within
the framework of FEWS NET, can become useful for understanding how trade flows change
during times of deficits in rainfall and food production.
Comparison of model predictions
To demonstrate that a country's NDVI anomaly over agriculture regions is really useful in the
model for improving price predictions in key markets, we compare predictions using models
specified using 2003-08 local maize prices, and evaluated the results using price data during
the 2009-12 period. Although the model has been implemented on all data, in Table 7.1 we
present the results of maize and millet price predictions from 42 markets in 11 food insecure
countries in Africa. In this analysis, a “naïve” model, comprising only previous-month (or
Markov)-based price forecasts, is compared to three other models:
As we add more parameters to the model, we expect to capture more variability in the price
dynamics and thus have a lower percent error in markets where these elements are important.
Table 7.1 has highlighted in grey the lowest root mean square percent error compared to the
actual prices during the 2009-12 period, which shows the model that is best able to capture
the observed variability of the prices in the market (Kshirsagar 2013).
In seven of the 42 markets the Markov-based predictions alone are best able to capture
price dynamics and have the lowest error as measured by RMSE. In 21 markets, the Markov
model and information on seasonality model is able to improve the forecast. In some markets,
this improvement is only marginal but in others the root mean square percent error (RMSPE)
between the predicted and observed prices declines by 2 to 4 percent. As discussed in Chapter
6 , seasonality is a critical element of not only local food production but also of food prices in
many regions where food security is a problem and the local markets do not function well.
Thus the impact of including a seasonal component to a Markov model demonstrated in this
analysis shows the importance of including information on seasonality in price modeling
When adding information on NDVI, the price prediction is improved in four markets -
Mbeye and Arusha in Tanzania, Garissa in Kenya and in N'Jamena Chad. These locations are
isolated from the world markets and from regional trading centers where the world market
prices are influential. When the NDVI and the world market prices are included together, a
further improvement in the prediction of prices in ten additional markets is seen. Some of
these predictions are significant improvements over the naïve Markov model, reducing the
errors from 24 to 16 percent error.
The standard set in this assessment is very strict, since in a model focused on providing actual
predictions in an operational framework, we would not train the model with prices that are out