MODELING THE IMPACT OF CLIMATE VARIABILITY ON LOCAL FOOD PRICES - Food Security, Food Prices and Climate Variability - page 142

Estimate price seasonality using trigonometric regressions

Select pixels that match seasonal price patterns:
- NDVI is falling and prices are falling
- NDVI is increasing and prices are increasing
- Both
Selection is based on 2003-08 monthly data

Select all NDVI pixels in the price location's region
(e.g., Eastern/Southern Africa) as candidate pixels
(exclude forests, water bodies)

Generate six different NDVI anomaly time series
based on matching criteria and country where the
pixels are located. Set anomalies to zero when NDVI
is typically falling

Use the NDVI anomaly signals as input
into a Hidden Markov Model that
assumes latent seasonality and
autoregressive (momentum) processes

Choose optimal model from a set of
models with different NDVI signals,
lag-lengths of NDVI and world price
shocks and different inputs
Model evaluation is based on out-of-
sample forecasts for 2009-11

Characteristics of
optimal model:
(i) set of informative NDVI
pixels
(ii) lag-lengths of NDVI/
world price impact

Output: map showing
candidate short-run
trade linkages

PRIMARY OUTPUT:
Price forecasts
(under alternative loss
functions)

Estimates of
impact of NDVI
anomalies and
world prices

Output 3: typology of
locations

FIGURE 7.5 Summary of methods for price-NDVI model.

The model then uses the seasonality of food prices together with NDVI to estimate which
NDVI time series is most related to food price variation for a particular model. The NDVI
pixels from agriculture regions in a country are averaged and used to create anomalies, which
can capture weather shocks on food prices. Then the model assumes seasonality and auto-
regressive momentum processes. Next, multiple models are developed with different NDVI
signals, lag-lengths and inputs and the algorithm chooses the optimal model from a set using
regression statistics. The optimal model is the one with the best out-of-sample performance
over three years (2009-11). The optimal model in one location could differ from the optimal
model in another location based on several criteria:

• Relevantpixelscouldbeexclusivelyinsideonecountryoralsoinothercountriesinthe
region. The pixels could be close or farther away from the location depending on market
size and trade dynamics.
• Theoptimalmodelcouldpotentiallyincludeworldpricesasanexplanatoryvariableor
it may not.
• NDVIanomaliesandworldpricescouldhavedifferentlag-lengths.
• Persistenceofmoisturedeicitsandpricemomentumfrompreviousyearscouldbe
incorporated into current year estimates.

Such a model would provide additional information about the location and the relationship
between price dynamics and weather shocks that is currently not available. The model then