Validation of operational seasonal rainfall forecast in Ethiopia


  • Diriba Korecha,

    Corresponding author
    1. National Meteorological Agency of Ethiopia, Addis Ababa, Ethiopia
    2. Geophysical Institute, University of Bergen, Bergen, Norway
    • Corresponding author: D. Korecha, National Meteorological Agency, PO Box 1090, Bole Road, Addis Ababa 1090, Ethiopia. (

    Search for more papers by this author
  • Asgeir Sorteberg

    1. Geophysical Institute, University of Bergen, Bergen, Norway
    2. Bjerknes Centre for Climate Research, University of Bergen, Bergen, Norway
    Search for more papers by this author


[1] Operational rainfall forecasts using the analog method have been issued in Ethiopia since 1987. We evaluate the performance of the forecast system for February–May and June–September rainy seasons over the period 1999–2011. Verification is performed using rainfall data obtained from Ethiopian meteorological stations covering eight homogeneous rainfall regions used in the forecasts. The results reveal that forecasts issued by the National Meteorological Agency (NMA) of Ethiopia, for the past 12 years have a weak positive skill for all eight regions compared with climatology. In terms of ranked probability skill scores, the values are all lower than 10% indicating that the forecast skill is modest. The results further suggest that the forecasting system has bias toward forecasting near-normal conditions and has problems in capturing below average events. In contrast, the forecast has some positive skills in ranking the wet years of February–May season, particularly over the regions where there is high seasonal rainfall variability with significantly positive rank correlations for the above average years. For the main season, however, the forecast is not able to rank wet years or dry years. The extreme low and high rainfall events are mostly missed by the forecast scheme. The results indicate rather low forecast skill for extreme rainfall events in both seasons. Generally, the results indicate that NMA's forecasts have low but positive skill as it is common with results from other forecasting systems for the Greater Horn of Africa region. The underforecasting of dry events is the most serious shortcoming of the system.

1. Introduction

[2] With irrigation covering only 1% of the soil that feeds more than 85 million people, the link between rainfall and agricultural yield is close in Ethiopia. It has been documented that food shortage and scarcity of water have led to local and nationwide famines, mainly due to the complete or partial failure of short and long rainy seasons over various parts of Ethiopia [e.g., NMSA, 1996]. The failure of seasonal rainfall is often caused by either misplacement or weakening of large-scale seasonal rain-producing systems. Attempts have been made to model these systems and factors out that could cause such failures in rainfall and numerous statistical and dynamical prediction models have been developed worldwide [Goddard et al., 2003; Barnston and Mason, 2011].

[3] Annual rainfall characteristics of Ethiopia are classified into three rainy seasons as documented by many authors [Gissila et al., 2004; Segele and Lamb, 2005; Korecha and Barnston, 2007]. These distinct seasons are; the dry (October–January), the small rainy (February–May), and the main rainy (June–September) seasons. The seasons are locally defined as Bega (October–January), Belg (February–May), and Kiremt (June–September). Although delineation of distinct regions and rainy seasons are difficult due to the complex topography of the country and high rainfall variability, the present forecast verification is done based on the existing homogeneous rainfall regimes.

[4] Variation of rainfall depends mainly on an advection of moist air and the location and intensity of rain bearing systems over the vicinity of Ethiopia. For instance, the westward propagation of weather disturbance developing over the Indian Ocean and Arabian Sea as well as southerly moisture flow are widely known rain-producing features for the eastern African subregion, including Ethiopia [e.g., Nicholson, 2000; Segele and Lamb, 2005]. In Kiremt, the rain-producing systems and their features are mostly associated with the establishment of synoptic and planetary scale systems such as Inter-tropical Convergence Zone (ITCZ), southwest monsoon components, and short-lived weather disturbances forming over the Arabian region [Camberlin, 1997; Segele et al., 2009a; Diro et al., 2011; Wolff et al., 2011].

[5] Sileshi and Demarée [1995] indicated that the Indian Ocean is one of the main moisture sources for Ethiopian rainfall. Similarly, Camberlin [1997] argued that the Kiremt rains (June–September) of Ethiopia rely on moisture advection from the Congo Basin through the southwesterly monsoon. Mohamed et al. [2005] also indicated that the oceanic sources of atmospheric moisture over the Nile basin are the Atlantic and the Indian Oceans. In contrast to the Kiremt rain, eastward traversing of midlatitude frontal systems often triggers unseasonal rain during reasonably dry seasons (October–April) over portions of northern, central, and eastern Ethiopia [Kassahun, 1987; Nicholson, 2000]. The skill of the predictability of seasonal rainfall therefore depends on an extent toward which the prediction systems could quantify the depth and the flow of moisture, regional, and global systems as well as the atmospheric dynamics that initiate seasonal rains.

[6] Statistical relationships between the Ethiopia rainfall and other meteorological parameters or sea surface temperatures have been investigated in a number of papers [e.g., Degefu, 1987; Segele and Lamb, 2005; Korecha and Barnston, 2007; Diro et al., 2011]. By assessing the lag-time correlations of SSTs with seasonal rainfalls of various regions over Ethiopia, the National Meteorological Agency (NMA) has issued seasonal forecasts three times a year since 1987 as documented by Korecha and Barnston [2007].

[7] The role of the El Niño Southern Oscillation (ENSO) on the Ethiopian seasonal rainfall is well documented and associated hazards often coincide with the occurrence of major ENSO events [NMSA, 1996; Camberlin, 1997; Bekele, 1997; Tsegay, 1998; Gissila et al., 2004; Segele and Lamb, 2005; Korecha and Barnston, 2007; Diro et al., 2011]. More recently, Araya and Stroosnijder [2011] documented how various ENSO events disturbed the onset and cessation of seasonal rainy season over northern Ethiopia. In association to mitigating river floods, Wang and Eltaher [1999] underscored the importance of ENSO information for forecasting precipitation over Ethiopia. Moreover, Block and Rajagopalan [2007] pointed out that ENSO phenomenon is the main driver of the interannual variability in seasonal precipitation in the Blue Nile basin, with El Niño (La Niña) events generally producing drier (wetter) than normal conditions. Furthermore, Elagib and Elhag [2011] provided evidence of an ENSO footprint on seasonal rains over about two-thirds of the area of the Sudan. It is, therefore, broadly argued that Ethiopian seasonal rainfall performance is strongly linked to ENSO.

[8] Since the beginning of seasonal climate prediction in Ethiopia, NMA has gone through continuous improvement in order to enhance the skill of predicting strong seasonal rainfall anomalies for various parts of the country. In recent years, research papers have proposed various statistical techniques to predict the major rainy season in Ethiopia [Gissila et al., 2004; Korecha and Barnston, 2007; Segele et al., 2009a, 2009b; Diro et al., 2009, 2011]. However, to what extent these forecast techniques are of better quality than the NMA forecasting system is unknown. This is due to the fact that few attempts have been made to assess the skill of the operational seasonal rainfall forecasts issued by NMA or other regional and international climate prediction centers for Ethiopia. To address this we here attempt to assess the skill of the NMA's operational seasonal predictions in order to provide a benchmark against which new climate prediction systems can be measured.

[9] The main objective of the present study is, therefore, to verify the skill of seasonal rainfall forecast that have been issued by NMA for the period 1999–2011 for the February–May and June–September rainy seasons. Although ONDJ (October–January) is the dry season over the Kiremt-rain-benefiting regions of Ethiopia, it may have better predictability [e.g., Indeje et al., 2000], it will not be considered in this study.

[10] The paper is arranged as follows: in section 2, the seasonal rainfall forecasting systems used by NMA are presented. The database (archive of forecasts and observations) and validation techniques are explained in section 3. Section 4 describes and discusses results. Conclusions and recommendations are given in section 5.

2. Seasonal Rainfall Forecasting System at National Meteorological Agency (NMA)

2.1. Background

[11] The seasonal forecasting systems and techniques used by NMA have been documented in several papers [e.g., Bekele, 1997; Korecha and Barnston, 2007; Diro et al., 2011]. In this section, some of the essential components of seasonal forecasting and procedures used by NMA are briefly described. As seasonal climate predictors, NMA uses indices of sea surface temperatures (SSTs) over the tropical Pacific Ocean, the Southern Oscillation Index (SOI), the Multivariate ENSO Index (MEI as described by Wolter and Timlin [1998]) and the ENSO (El Niño-La Niña) outlook obtained from NOAA/CPC. Historical and current Niño 3.4 SSTs (the Niño 3.4 region is located in the central equatorial tropical Pacific Ocean) are used to select years with ENSO evolution similar to the current year. Rainfall prediction for the current year is then based on rainfall observed in these analog years. Monthly SSTs are compared for several months in advance of the season to be predicted (Figures 1 and 2). For example, in order to predict rainfall of the June–September (JJAS) season, Niño 3.4 SSTs for January–May of the current year are compared with SSTs for the same months in 1970, 1971, etc., and analogs (years with similar ENSO evolution) are identified. By considering the current and future ENSO states, the best three analog years are selected from the primarily listed similar years (Figure 2). This procedure is done using graphical and rank correlation techniques. Following these steps, the seasonal rainfall of each station is calculated for each analog year that the station rainfall in each analog year is expressed as a percentile of the full climatology using a percentile statistical approach. Station-based seasonal rainfall percentiles [following Gibbs and Maher, 1967] are then used to calculate tercile categories (0–33; 34–66, and 67–100%) for each homogeneous rainfall region. NMA's seasonal rainfall forecast is then prepared as a probability of the regional seasonal rainfall being below, near, and above the climatological normal (in this case, the mean from 1970 to the year under consideration). The tercile rainfall categories, which are more commonly known as the probabilities, refer to the likelihood that the region-averaged rainfall will be below, near, or above average as the anomalies in seasonal (4 month) rainfall are often large in geographical scale. This forecast format is motivated by the simplicity of the forecast presentation and is used by many operational seasonal forecast centers. Figure 2 (top right plot) shows an example of an official NMA rainfall outlook for the JJAS 2010 rainy season. Finally, NMA issues the seasonal rainfall forecast for each season (FMAM, JJAS, and ONDJ), 1–2 weeks prior to the normal onset date of each season (Figure 1).

Figure 1.

Seasonal rainfall forecasting system of the National Meteorological Agency of Ethiopia.

Figure 2.

Schematic diagram showing analytical steps in the preparation of seasonal rainfall forecast by the National Meteorological Agency of Ethiopia.

2.2. Justification for Analog Forecasting Method

[12] The use of analog methods to generate climate forecasts is an attractive idea, not the least because of its conceptual simplicity, and many meteorological institutions around the world either still use the analog forecasts or have done until recently [WMO, 2002]. The analog methods used to forecast rainfall anomalies directly, or through the intermediary of an anomaly flow pattern. In some cases, they are used more indirectly, e.g., the Australian Bureau of Meteorology determines analog years of the Southern Oscillation Index (SOI) as a first step to subsequent analyses [Drosdowsky, 1994]. The techniques involve searching of the historical data, identifying previous periods that resembled the immediate past period, and predicting the following season's rainfall anomalies on the basis of what happened on those previous occasions. Analog forecasting techniques have been used in climate forecasting for a long time. Namias [1968] reviews the early history, and Nicholls [1980] presents a somewhat more recent view [Brett and Thompson, 2006]. There were also revivals of interest in the use of analog forecasting techniques at the end of the 1980s, particularly in the United States and New Zealand, with the papers [e.g., Barnston and Livezey, 1989; Chapman and Walsh, 1991; Livezey et al., 1994; Brett and Thompson, 2006]. In their comparative study, Barnett and Preisendorfer [1978] found that the use of climate systems evolution in defining an analog sometimes gave a superior prediction and at other seasons and lead times gave a worse result.

[13] In identifying the predictors for Ethiopia rainy seasons, previous researches guide the selection of the most appropriate predictors from the historical archives. In this regard, a number of observational studies have identified the use of Equatorial Eastern Pacific Ocean SSTs as potential predictors of Ethiopia rainfall anomalies with some lead-time in advance [Korecha and Barnston, 2007]. For instance, during the El Niño/La Niña events, Ethiopia experiences less/more rainfall in the northern half and more/less in the south and southeast regions during the Kiremt season. Shanko and Camberin [1998] have found that Indian Ocean sea surface temperatures have an important influence on Ethiopia seasonal rainfall. They have found that higher SSTs in the Eastern Indian Ocean, for instance, generate a lot of tropical cyclones, which are resulted in drier conditions in the north, east, and south of the country during Bega and Belg seasons.

[14] Various teleconnection patterns are linked to Indian, Atlantic, and Pacific Oceans, where they produce different climatic anomalies in various parts of Ethiopia [Segele and Lamb, 2005; Diro et al., 2011]. Thus, when predicting Ethiopia seasonal rains, it is important to allow for seasonality and regional rainfall feedback to various teleconnection indices. However, ENSO-indices have well been identified as the potential preseason indicators and thus became the basis for the analog forecasting techniques in Ethiopia [Korecha and Barnston, 2007]. ENSO indices are being retained year round, but allowing these indices to be weighted differently from season to season as well as from region to region, depending on the direct linkage between regional rainfall pattern and SST anomalies. The analog forecast methodology is therefore now run operationally at the National Meteorological Agency. As part of NMA's long years' early warning program on the monitoring of climate variability, national seasonal rainfall outlook forums usually convene three times a year to discuss seasonal climate anomalies over Ethiopia since 1996. A range of guidance material is used, and the analog seasonal climate prediction method that identifies 3–5 analog years is a very useful prediction tools in providing tercile rainfall probabilities for each season in Ethiopia.

[15] The National Meteorological Agency of Ethiopia has therefore integrated an analog forecasting technique in its seasonal climate prediction system. The major significant analog forecasting technique currently used in NMA is its dependence on the scientific innovations and explorations of ENSO. The technique has improved seasonal rainfall predictions and has better consideration of oceanic longer time memory of SST anomalies, and has been used in the countries where the computing facility is very weak. It is also widely used in the tropical regions as the predictability skill of seasonal rain is relatively dependable. Hence, examining of national rainfall anomalies on the basis of ENSO-teleconnection, as well as the extent of the extremity of droughts and flooding can be resolved. This can give a unique opportunity for NMA to provide timely early warnings on the adverse effect of climatic anomalies within the reasonable lead time. This is unique and to the best of our knowledge, no other climate prediction technique has used such simple, less expensive computer facility and makes use of few indices because it is too expensive to run the state-of-art of modern general circulation models. So, this guarantees NMA's seasonal climate prediction technique is well maintained and acquired the modest capacity in capturing the drier and wetter occasions without using any other advanced climate prediction models. Thus, this is the superior point of the current analog forecasting technique of NMA.

[16] NMA has divided Ethiopia into eight homogeneous rainfall regions. The classification is based on; typical rain-producing systems affecting the region and spatial and temporal response of respective region to major atmospheric and oceanic circulation systems. Although some authors [Gissila et al., 2004; Diro et al., 2008, 2009] have proposed modifications to the NMA homogeneous rainfall regions, NMA still uses the originally defined eight rainfall zones for the preparation of seasonal rainfall forecast (Figure 3). Although, NMA has been issuing the seasonal forecasts for many years, the overall statistical performance of these forecasts has not yet been comprehensively documented. Bekele [1997] made qualitative forecast verification on the seasonal rainfall forecasts issued for the period 1987–1996 and claimed a seasonal rainfall forecast percent correct score of 75% or more. From the qualitative forecast assessment, we noted that underforecast of severe dry events may be a result of the fact that there is a greater reluctance to assign high probabilities for below average than for above average rainfall since in many parts of the country a warning of dry conditions would be considered more serious than wet conditions. In this paper, we revisit assessment of the forecasts using an objective verification approach.

Figure 3.

Homogeneous rainfall regions currently used for the preparation of seasonal rainfall forecast in Ethiopia. Meteorological stations used in this study are marked as “*.”

3. Database and Verification Technique

3.1. Database

[17] This verification is made for two rainy seasons over Ethiopia; February to May and June to September for the period 1999–2011. Monthly rainfall data from NMA meteorological stations (Figure 3) are used. Numbers of meteorological stations used in this study varies between 115 and 226. The period 1970–1998/99 is regarded as base period against which the observed seasonal rainfall in each verification year is compared. Missing data for any months are excluded from the verification analysis so as to avoid artificial data filling for the season under consideration. Seasonal rainfall forecasts that are available only in tercile rainfall probabilities maps (Figures 4a and 4b) obtained from NMA's seasonal climate prediction records. The spatial delineation of zones with the same set of forecast probabilities varied from year to year. The forecast maps have therefore been recast using the NMA's eight homogeneous rainfall regions (Figure 3), employing the method described in the next section. Meteorological stations used for the forecast verification are also organized in their respective homogeneous regions. To indicate the state of ENSO, we use the ENSO indices from NOAA/CPC data set as described by Korecha and Barnston [2007].

Figure 4.

Examples of tercile observed and probability forecasts for seasonal rainfall over Ethiopia. The two paired maps show forecasted (observed) rainfall probabilities for (a) FMAM and (b) JJAS rainy seasons. Here, the maps are presented as prototype. The seasonal rainfall forecast issued for the period 1999–2011, which did not display here have similar configuration.

3.2. Verification Technique

[18] In the preparation of seasonal rainfall forecasts, NMA uses eight homogeneous rainfall regions (Figure 3). The tercile rainfall probabilities for the eight regions are however, as we can see from Figures 4a and 5c often merged into fewer regions in the presentation of the forecast. In order to verify the forecast we split the merged regions into the eight original homogeneous rainfall regions based on Figure 3. This is achieved by superimposing Figure 3 on the forecasts maps for each year. When two or more forecast zones occupy one homogeneous region, the forecast probabilities for the zone that covered substantial portions of the region are assigned to the whole of the homogeneous region. This is done for the periods 1999–2010 for the June–September main rainy season and for 2000–2011 for the small February–May rainy season. Meteorological stations used for this verification processes (see section 3.1) are then grouped into their respective region. Seasonal rainfall totals for each station are then ranked in comparison to the values in the base years. For example, JJAS rainfall totals recorded at station “X” in 1999 is ranked within 1970–1999 rainfall time series and its percentile rank is assigned, accordingly. Similarly, the seasonal rainfall time series for 1970–2000 are ranked and percentile rank is assigned for the year 2000, and so on until all years to be verified scheme are ranked according to the rainfall magnitude. This is done for all stations within the region. The tercile observed frequency of occurrence for a given season are then calculated based on the number of stations having its seasonal rainfall ranked in the upper third (above normal), lower third (below normal), or in between (near normal) for the given year.

Figure 5.

Comparison of observed and forecasted the tercile rainfall probabilities over the two homogeneous rainfall regions in Ethiopia. Below, near, and above average rainfall probabilities are paired in the diagrams in order to identify the discrepancy between observed and forecasted rainfall during February–May season.

[19] Various verification techniques are described in the literature [e.g., Murphy, 1988; Jolliffe and Stephenson, 2003; Goddard et al., 2003; Barnston et al., 2010; Barnston and Mason, 2011]. Based on the format of our data set, we employ the following verification techniques to examine the bias, association, accuracy, and skill of the forecasting systems.

[20] First, we use a diagram presentation to compare the forecasted and observed seasonal rainfall probabilities for each tercile category. The diagram indicates how well the predicted probabilities of an event correspond to their observed probability for each category. The measure does not say if the seasonal predicted rainfall is strongly deviated to other tercile categories.

[21] Second, in order to assess any directional bias, which is a systematic tendency to assign too much or too little probabilities to particular tercile categories, we computed the directional bias (DB) as

display math

where fk and ok are the average forecasted and observed seasonal rainfall probabilities for the years 1999–2011 for station k in each region, respectively. The three tercile categories used in this particular case are below, near, and above normal rainfall probabilities. If there is no directional bias the result is always zero. In contrast, if the forecast probabilities are too high, DB will be negative and vise versa.

[22] Third, the spearman rank correlation test [Jolliffe and Stephenson, 2003] was applied to measure the statistical association between the forecasted and observed relative frequencies of rainfall categories in each tercile. Spearman correlation coefficient (SRC) is defined as the Pearson correlation coefficient between the ranked variables:

display math

where Di represents the difference between ranks of pair of data values for n observations. More specifically, Di is the difference between the highest rainfall tercile probability assigned for the forecast and the corresponding actual rainfall percentile for the ith year. The higher/lower value of SRC (approaches to 1) indicates if the forecast is able to rank the years within a tercile correctly. For example, if the forecast is able to assign high probabilities of a wet season to a year that was extremely wet and a lower probability of a wet season to a year that was less wet, the forecast will have good skill.

[23] We also evaluate the skill of the forecast using the forecast hit as described by Mason and Weigel [2009]. A strong rainfall forecast hit occurs when the tercile with the highest probability forecasted coincides with the highest observed probability. For example, if the probability of a wet season is higher than for both a normal and a dry season in both the forecast and observations it is counted as a hit. The skill of the forecast may be evaluated by how much larger the hit score is compared to the hit score that may be expected by chance (0.33).

[24] In addition to the above verification measures, we also use the ranked probability skill scores (RPSS) to the three forecast categories collectively [Goddard et al., 2003]. RPSS computes the relative skill of the probabilistic forecast over that of climatology, in terms of the forecast ability to assign high probabilities to the actual outcome and is defined as the difference in ranked probability score between the forecast and a chosen reference forecast [Goddard et al., 2003; Wilks, 2006; Barnston et al., 2010]. Thus, the RPSS measures the improvement of the multicategory probabilistic forecast relative to a reference forecast (usually the long term or sample climatology). It is similar to the 2-category Brier skill score, in that it takes climatological frequency into account. When RPSS is computed, the probabilities of the three forecast categories; below, near, and above averages are arranged in ascending order. The ranked probability score (RPS) is then calculated

display math

where ok is an indicator which is 1 if the forecasted and observed category coincided (for example, both have below average rainfall as the most probable category) or 0 otherwise. fk is the predicted probability in forecast category k (for k = 1, 2, or 3) for each station and forecast year, and N is the number of forecast categories (in this case N = 3). Low RPS indicates high skill, and vise versa. The RPSS is thus calculated as

display math

where RPSr represents the RPS value obtained from climatological forecasts. In our case, climatological value is 0.33 (any of the three terciles; below, normal, and above normal are equally likely to occur).

[25] In addition to using climatology as a reference, also we use ENSO as a reference to see if the forecast beats a pure ENSO-based forecast. The way this is done is that rainfall recorded at each meteorological station was ranked within 1970–2011. Then based on ENSO phases (El Niño, Neutral, and La Niña) numbers of stations within a region were stratified into terciles (below, near, and above average). For the 41 years of JJAS (1970–2010), 10(9) years are classified as El Niño (La Niña), and 22 years as neutral. Similarly, for 42 years of FMAM (1970–2011), 6(10) years are classified as El Niño (La Niña), and 26 years as neutral.

4. Results and Discussion

4.1. Categorical Forecast Skill

[26] The seasonal forecasts were evaluated to see if there is any directional bias. It can be seen from Table 1a that the forecast underpredict below average rainfall in all regions in the FMAM season. The underforecasted value ranges from 27 to 45%. The same tendency for underforecasting dry conditions is seen in JJAS (Table 1b). In contrast, the forecast was substantially biased toward the near average category in all regions both in FMAM and JJAS (Tables 1a and 1b). In general, above average rainfall occurred on average 28% (not shown in Tables 1a and 1b) of the cases during period 1999–2011 as compared to the reference climatological base period (1971–1998). Whilst the near average forecast probabilities remained above 45% exceeding the actual “probabilities” (Tables 1a and 1b). The bias of the prediction system toward near average indicates that lack of forecast sharpness in predicting events deviating from the normal.

Table 1. The Bias in Forecast Probabilities for Three Tercile Categories: (a) FMAM Rainfall Forecasts and (b) JJAS Rainfall Forecastsa
(a) FMAM Rainfall Forecasts
Rainfall RegionDirectional Skill of the Seasonal Forecast (%)
Below AverageNear AverageAbove Average
(b) JJAS Rainfall Forecasts
Rainfall RegionDirectional Skill of Seasonal Forecast (Forecasted/Observed)% − 100
Below AverageNear AverageAbove Average
  1. a

    Numbers with symbol “↓”indicate that the forecast system was under forecasted of observed rainfall is being underforecasted by the forecast, while “↑” indicates overforecasting.


[27] To further examine the skill of the seasonal forecasts, the probabilistic value assigned for each category is plotted with the observed rainfall percentile in Figures 5 and 6 (only shown for Regions I and II). Each paired bar diagrams show the comparison of the forecasted versus observed rainfall percentiles for each tercile categories. The near average rainfall probabilities were forecasted to be the most probable event in the order of 40–50% of the time (Figure 5), while the below and above average categories were forecast less frequently (22–33% and 24–30% of the time, respectively). From Figures 5 and 6, we observe that large numbers of stations with the lower tercile are not forecasted by the forecast system correctly. For instance, there is a large departure between forecasted and observed rainfall in 2002 and 2009, when many regions experienced deficient rainfalls. On average, the below normal rainfall probability forecast was the highest in 46% (ranging from 40 to 70% depending on the region) of the observed below normal rainfall events, while below normal rainfall probability was wrongly assigned as the most probable in 54% of the events. NMA's forecasting system sometimes forecasts the above average category as most probable for low rainfall years. This underforecast of severe dry events may be a result of the fact that there is a greater reluctance to assign high probabilities for below average than for above average rainfall since in many parts of the country a warning of dry conditions would be considered more serious than wet conditions.

Figure 6.

Comparison of observed and forecasted the tercile rainfall probabilities over the two homogeneous rainfall regions in Ethiopia. Below, near, and above average rainfall probabilities are paired in the diagrams in order to identify the discrepancy between observed and forecasted rainfall during June–September rainy season.

[28] Above normal rainfall probability forecast was the highest in 52% (ranging from 30 to 60% depending on the region) of the observed above normal rainfall events with a false alarm rate of 41% (Figure 6). In particular, the wet events of 2003, 2006, and 2007 were not predicted correctly. The extreme low and high rainfall events are mostly missed by the forecast scheme. Hence, the results indicate that low forecast skills were attained for strong rainfall events both for the two seasons.

4.2. Spatial Forecast Skill

[29] Forecast biases, standard deviations of the forecast and observed rainfall probabilities and the rank probability scores skill (RPSS) were computed for each homogeneous rainfall region in Ethiopia for the FMAM and JJAS rainy seasons between 1999 and 2011 (Tables 2a and 2b). The standard deviations show that strong interannual rainfall variability is marked, with higher standard deviations of the probabilities of the observed rainfall over each region much higher than the standard deviations of the forecasts probabilities (Tables 2a and 2b). This indicates that the forecasts issued for the seasonal rainfall varied much less than the actual rainfall.

Table 2. Statistical Values for Skill of Ethiopia Seasonal Rainfall Forecasts Over the Homogeneous Rainfall Regions During 1999–2011 for (a) FMAM and (b) JJAS Seasonsa
(a) FMAM Season
RegionRainfall ProbabilityMeasure of Discrepancy
(b) JJAS Season
RegionRainfall ProbabilityMeasure of Discrepancy
  1. a

    Forecast probabilities issued for each tercile category are verified against the observed relative frequency at each region. Biases, the difference between the average forecasted and observed probabilities are computed along with the standard deviations for observed (SDo) and forecasted (SDf) and are also included. RPSS are computed based on both individual station and regional rainfall performance for each homogeneous rainfall region.


[30] The aggregated RPSS for each homogeneous rainfall region shows positive, but low predictability skills. The RPSS values for FMAM are slightly higher than for the JJAS season mainly over the Belg rain-benefiting regions (Table 2). This indicates that the short rainy season has been slightly better predicted than the main rainy season over the regions that experience bi-modal rain type. But the difference is only 0.09, which is unlikely to be significant. To investigate whether the forecasting system has skills in all homogeneous rainfall regions, the hit rate of the seasonal rainfall forecast over each region is shown in Figure 7. As we noticed in the previous sections, the forecast skill varied from region to region in addition to its seasonal variation. Among the eight homogeneous rainfall regions, the forecast is above 0.33 only for three of them (eastern parts of the country) during the FMAM season. For the JJAS season, the forecast system exceeds the climatological chance (0.33) in four of the eight homogeneous regions (two regions in the west and two in the south).

Figure 7.

Seasonal forecast hit rate, for FMAM and JJAS seasons, scores probabilities given for an event mostly occurred for below, near, and above average rainfall category. Horizontal black line represents the forecast skill obtained by chance or climatology. The skill of a forecast is evaluated based on distance between the hit and the level that is reached by chance (33%).

[31] The relative skills of the probabilistic forecasts were assessed over that of climatology and ENSO RPSS are then calculated and presented in Table 2 for FMAM and JJAS. Table 2 shows the forecast to have slightly better skill than climatology with RPSS values up to 8–9% in a few regions during the FMAM season over the regions experiencing bi-modal rain types, while in the case of JJAS the RPSS is somewhat lower (4–6% in five of the eight regions). Although the RPSS indices are weak, they are all positive, indicating the presence of some predictability skill for both seasons over Ethiopia.

[32] Figures 8a and 8d show the spatial RPSS patterns with climatology as the reference in both rainy seasons. Spatial RPSS patterns indicate that the forecast system performs better than climatology in much of the country, however the values are low. The FMAM season (Figure 8a) has the highest forecast skill with values above 10% over south Ethiopia (Figure 8a). In contrast, for the JJAS season the RPSS was worse than climatology over the southwestern lowlands and eastern portions of the country (Figure 8d). Overall, in many regions the forecast skills are slightly higher over the Belg-rain-benefiting regions in FMAM compared to JJAS. This is possibly related to the persistent nature of rainfall producing systems and their strong spatial variability during the spring season. In contrast, JJAS rainfall is more predictable with relatively higher RPSS over the Kiremt is mainly the main rainy season.

Figure 8.

Spatial distribution of RPSS averaged for the study periods. RPSS are computed for (a–c) FMAM and (d–f) JJAS seasons separately based on point meteorological stations. Three RPSS are analyzed in order to evaluate the skills of NMA'S seasonal forecast versus climatological reference and ENSO climatology (Figures 8b and 8d). Also ENSO climatology is compared against climatological references (Figures 8c and 8f).

[33] Figures 8c and 8f show the maps of RPSS for ENSO climatology and climatological references. When RPSS of ENSO climatology is compared with the forecast issued by chance (assigning equal chances for the three tercile rainfall categories), ENSO information alone can indicate the direction of the seasonal rainfall anomalies particularly during JJAS season for northern Ethiopia. For this study, ENSO climatology is computed based on the Oceanic Nino Index of NOAA as documented in NOAA [2013]. The results indicate that injection of more weight of ENSO information into the seasonal predictability scheme would improve the forecast skill in parts of Ethiopia during the rainy seasons, a conclusion also drawn by Korecha and Barnston [2007].

[34] As the above analysis indicated, the ENSO information alone provided some skills in predicting seasonal rains. This would, therefore, further lead us to investigate the quality of the seasonal forecast by considering ENSO climatology. Results from Figure 8b show that in FMAM the forecasting system have better skills in regions I, III, and VII, while ENSO climatology was better in regions IV, VI, and V (Figure 8b). In JJAS, the ENSO climatology outperforms the NMA forecast over the major portions of the country (Figure 8e) with the exception of the dry south-eastern parts. The underperformance of the seasonal forecast could emerge probably due to the fact that either the forecast system has given too little weight to the ENSO cycle or it has underestimated the ENSO impact over the aforementioned regions.

[35] Overall, the RPSS results indicate that NMA's seasonal rainfall forecasts have modest positive skill compared to climatology. In contrast, when the forecast is compared to ENSO climatology, it performs poorly, particularly over the central and northern regions in JJAS.

4.3. Variation of Forecast Skills With Seasons

[36] In order to assess if the forecast is cable to rank the years within a tercile correctly, Tables 3a and 3b show the Spearman rank correlations between the observed and forecasted tercile categorical seasonal rains. Eight of the 48 forecast series considered (eight regions, three categories, and two seasons) have a significant positive correlation. Most of the significant values were noticed for the above average rainfall forecasts during FMAM, which have statistically significant values for the regions in the south (regions III, VI, VII, and VIII, Figure 3). In general, the correlations in the JJAS season (Table 3) are weaker than in FMAM. The weak capabilities of the forecast to be able to rank the years within a tercile correctly were partly due to the fact that the system is biased toward the near normal rainfall category.

Table 3. The Association Between Forecasted and Observed (a) FMAM and (b) JJAS Seasonal Rainfall Probability for Each Forecast Categorya
(a) FMAM Season
Rainfall RegionCorrelation Between Observed and Forecasted Rainfall Probability
Below AverageNear AverageAbove Average
(b) JJAS Season
Rainfall RegionCorrelation Between Observed and Forecasted Rainfall Probability
Below AverageNear AverageAbove Average
  1. a

    Correlations are computed using Spearman rank correlations inline image. Values indicated as * and ** are statistically significant at 90 and 95% probability levels, respectively.


[37] The frequency of stations having seasonal rainfall of below, near, and above normal categories each year on a national scale is presented in Figure 9. It should be noted that the national rainfall index is biased toward regions with a dense station network. For the FMAM season 2000, 2002–2004, 2008–2009, and 2011 were severe drought years (Figure 9a), in line with the analysis of Viste and Sorteberg [2012]. In contrast, Figure 9b shows JJAS rainfall performance in tercile categories. Unlike, FMAM which experienced a higher number of years with deficient rains, the JJAS seasons over the period 1999–2010 seemed relatively stable; severe droughts occurred only in 2002 and 2009, when more than 60% of meteorological stations recorded below average rainfall.

Figure 9.

Nationalized tercile seasonal rainfall categories for: (a) FMAM and (b) JJAS rainy seasons. The dash line represents the climatological buffer zone (33.3%), which divides the degree of dryness or wetness of each season on the national scale.

[38] Yearly national RPSS values for the period 1999–2011 are calculated for the FMAM and JJAS seasons (Figure 10). The values are computed by averaging RPSS of each station (thus, it is bias to the regions with many stations). The results showed that the forecast system has positive skill on a national level except for the dry FMAM season of 2002 and 2009 and during JJAS 2000, 2004, 2005, and 2010. The highest skills (above 10%) are found in 2005 and 2011 for the FMAM season and during 2002 and 2009 for JJAS. With few exceptions the nationally averaged RPSS scores are slightly positive.

Figure 10.

Mean yearly RPSS values for FMAM and JJAS seasons for Ethiopia. Negative (positive) values indicate poor (good) forecast skill.

4.4. Reproducibility of Analog Climate Prediction Method

[39] The techniques of using the historical analogs to formulate a forecast, known as analog prediction method has been used to varying degree of success by different researchers [Agarwal et al., 2001]. Agarwal et al. [2001] further stated that this method is based on the premise that interseasonal changes in the climate system occur similarly from one instance to another, such that when the system is in the same state it was for the same season in some past years, a sequences of events similar to those which occurred in the past instance may be expected now also. For instance, Penland and Sardeshmuckh [1995] have reasoned that SST alone contains all of the relevant dynamics to a large extent, and thus there is sufficient ground to use SST observations for making predictions for up to nine months. Besides, conceptual simplicity, the main advantage of empirical prediction methods like analog techniques is that predictions are made using observed data alone, whereas in the case of coupled models the predicted SST fields may be affected by the way different physical process (e.g., air-sea coupling) have been parameterized in the model [Agarwal et al., 2001].

[40] Hammer et al. [2000] stated that the more historical data available the better the analog years could be identified that enable to capture possible climatic fluctuations and direction of change in climate variables. Development in good practices of climatic factors would also facilitate inclusion of effects of potential climate variation, which may underline trends in historical data that need to be considered when using historical analogs. The point remains, however, that of appropriate methods to derive analog years in order to connect forecasts and applications need to be viewed as an essential component of forecasting research and development [Hammer et al., 2000]. The issuing of forecast as simple probability statements is better, but provides only general information. In this regard, defining analog years enable the policy makers to capture climatic variability in a way that could formulate riskiness of alternatives decisions and to be evaluated by examining each year in the analog set separately.

[41] Hammer et al. [2000] further pointed out, averages are often a far less meaningful static of the probability distribution of outcomes than some consideration of the likelihood of exceeding some critical system state be it profit or land condition. Hammer et al. [2000] also noted that trends and phases of the ENSO have also been employed to provide seasonal climate outlook maps and tercile rainfall probabilities, from which forecast maps that describe the chances of rainfall in the above, average, or below average can be generated. With this understanding, the analog prediction method that has been verified in the case of Ethiopia contains a lot of scientific and application merits. In previous sections, it has been well articulated that in order to predict anomalies of rainfall amounts for the future season, analog method searches for a similar time evolution of seasonal rainfall in the historical data set.

[42] Pacific warm (El Niño) and cold (La Niña) episodes based on a threshold of ±0.5°C for the Oceanic Nino Index (ONI as computed based on 3 month running mean of ERSST.v3b SST anomalies in the Nino 3.4 region (5N–5S, 120–170W) have been used as the main database in order to identify more resemblance analog years with the on-going features from the historical episodes or nonepisodes [NOAA, 2013]. To demonstrate the reproducibility of analog prediction method, we use 760 grid box from the gridded rainfall data set that have been generated by blending observed rainfall data from coarse stations with satellite rainfall estimates [Dinku et al., 2011; NMA, 2013]. Gridded data set has records back to 1983. We assemble gridded rainfall data into each homogeneous rainfall zone and compute regional rainfall totals for the period 1983–2010. The linear correlation coefficient is then computed for each homogeneous rainfall region based on rainfall total and ONI for simultaneous and lags seasons. In order to investigate the response of regional rainfall to ENSO episodes, the correlation values are plotted against the 3 month running season (Figure 11). For the sake of clarity, the results reveal that there is a strong linkage between seasonal rainfall and ENSO episodes. We therefore strongly argued that ENSO-based analog years selection is scientifically sound and can further be explored as well as adopted for the tropical regions. Figures 11a and 11b that the analog methods are reasonably skilful in indicating the direction of seasonal rainfall conditions for both rainy seasons in Ethiopia. One of the limitations of analog prediction method, however, is in fact the shortening of lag time relationship exists between ENSO and Ethiopian rainfall. It can be expected that the accuracy of the present prediction method will improve as other global indices include in the selection of analog years with time.

Figure 11.

Correlation between seasonal rainfall totals of each rainfall regime and Nino 3.4 Sea Surface Temperature (SST) as computed for Ocean Niño Index (ONI). ONI is computed as an overlapping of consecutive months (e.g., JJA means June–August) as documented by NOAA [2013]. The linear correlation coefficient is computed between seasonal ((a) FMAM and (b) JJAS) rainfall totals (NMA, 2013) and ONI for the same season and preceding months.

[43] Generally, the analog prediction technique employed in Ethiopia has been compared with the existing regional and international model products (not shown here). The results revealed that the present method has the potential to predict most of anomalous years during February–May and June–September rainy seasons, specifically over the regions where rainfall is typically deficient. This argument could be further substantiated by considering simultaneous relationship that exists between February–May and June–September seasons. It is depicted that ENSO episodes enhance/suppress February–May/June–September (Figure 11). For instance, with the exception of northwest (Regions II, RII) and southwest Ethiopia (Regions III, RIII), El Nino usually increases the likely of above average seasonal rainfall during February–May (Figure 11a). In contrast, El Nino increases the likelihood of below average rainfall during June–September season over the major portions of Ethiopia (Figure 11b). As far as the scientific merit of analog prediction method is concerned, this method has potential particularly over the tropical regions where ENSO is strongly teleconnected with seasonal rainfall feature. This is clearly evidenced from our analysis as depicted in Figure 11. The method is therefore, highly recommended and reproducible for the regions that have similar rainfall pattern like Ethiopia, specifically where the climate models suffers a lot for its poor performance in capturing seasonal rainfall variation. It seems fairly certain that the method will be superior in the field of seasonal climate prediction by providing more reliable forecast even in the arena of newly emerging complex computational technologies. Because it is very impractical to run expensive climate models with the existing limited computing facilities and expertise in Ethiopia.

5. Conclusions and Recommendations

[44] In the present study, we evaluated the skill of the National Meteorological Agency of Ethiopia's operational seasonal rainfall forecast for the February–May (FMAM) and June–September (JJAS) rainy seasons for the period 1999–2011. Our analysis shows that the forecasting system is biased toward the near-normal category. The hit rate for forecasting the correct tercile is above 33.3% (the value that may be obtained by chance) for 8 out of 16 forecasts series. The ranked probability skill scores (RPSS) which computes the relative skill of the probabilistic forecast over that of the climatology is positive for all 16 forecasts series, indicating that the forecast has skill compared to chance. However, the values are all lower than 10% thus the forecasts skill generally ranges from weak to moderate, depending on the season and regions under question.

[45] The results further suggest that the forecasting system has problems in capturing below normal rainfall events. This is particularly pronounced during the February–May rainy season. This underforecasting of dry events is of great practical importance. The forecast showed slightly higher skills for above than below normal rainfall categories during both seasons and indicate that there is a greater reluctance to assign higher terciles for below normal than for above normal rainfall as a forecast for dry conditions would be considered more serious and may lead to initiation of drought preventive actions.

[46] The forecast has some skill in ranking the wet years of the FMAM season, where four of the eight regions have significantly positive rank correlations for the above normal years. In the JJAS season, the forecast is not capable of ranking neither the wet nor dry years. The seasonal difference in skill found in this study was also noted by Batte and Deque [2011]. The predictability skill results computed over the Greater Horn of Africa varied between the MAM (March–May) and SON (September–October) seasons, and suggested that seasonal rainfall predictability was higher for SON than MAM [Batte and Deque, 2011].

[47] In the above validation, it may seem that the forecast is not performing too well, but is not worse than other seasonal forecast attempts. The RPSS values that are computed for NMA's forecast systems seem to be comparable to the values computed for IRI's prediction scheme as documented by Goddard et al. [2003] and Barnston et al. [2010], the CFS seasonal forecast [Sooraj et al., 2012], the ENSEMBLES project [Batte and Deque, 2011], and African RCOF forecasts [Mason and Chidazambwa, 2008].

[48] We have already pointed to the importance of ENSO as potential indicator for the Ethiopian rainfall. When RPSS of ENSO climatology is compared with the forecast issued by chance and ENSO information alone had some skill in indicating the direction for the seasonal rainfall anomalies particularly during JJAS season for the northern Ethiopia. The results show that more weight on ENSO information into the seasonal predictability scheme would improve the forecast skill for JJAS rainy season. However, the information from ENSO alone is limited, particularly, due to the seasonality of ENSO and its predictability barrier during the northern hemisphere spring [Webster and Hoyos, 2010]. In order to make considerable improvements in the forecast some of the underlying factors other than ENSO are needed to be identified. In this case, Ng'ongolo and Smyshlyaev [2010] have shown that the phase of the QBO prior to the East African March–May (MAM) seasonal rainfall is a useful predictor for the seasonal rainfall. They argued that this finding is useful particularly in the case of the East African long rains, which is FMAM in this case, for which ENSO provides only limited skill.

[49] We suggest that NMA should work further to make appropriate improvement on the predictability of seasonal rainfall systems, especially for below normal rainfall categories. From a practical use, quantifying and providing accurate forecast for this category would be very beneficial for the user communities. The present RPSS analysis shows that the predictability skill for the June–September rainy season is poor. Therefore, work on identifying the underlying rain-producing systems and examine closely their physical linkage with larger-scale surface indices such as ENSO or circulation indices (QBO) which have been shown to have some predictability skill on the seasonal scale should be conducted. Moreover, the selection of homogeneous rainfall regions is important as this is the spatial scale for which the forecast is issued for. Merging heterogeneous rainfall regions into one region may also distort the level of seasonal forecasting skill over various parts of Ethiopia. In this regard, further research on how to separate the country into useful rainfall regions may be beneficial for the forecast quality.

[50] Major advantages of NMA's ENSO-based forecasting technique over the other prediction methods are that it automatically finds out closely matching patterns from the corresponding historical occasions. This feature considerably minimizes data processing requirements. The skills demonstrated with this fairly simple method are high and have immense potential for practical purposes. It is argue that the on-going statistical and dynamical climate prediction models will improve with time with the wealth of understanding of the climatic factors and would be able to simulate and produce skilful extended-range forecasts of the Ethiopian intraseasonal rainfall variability. For the time being, however, a judicious and practical way is to use the existing NMA's ENSO-based analog methods for seasonal predictions, with reasonably a few weeks in advance.

[51] Finally, in addition to relying on the existing analog method NMA should explore the possibility to improve the forecast by using other dynamical and statistical forecast techniques that uses the seasonal forecast information from available global modeling systems.


[52] This work was carried out with the support from the Ethiopian Malaria Prediction System (EMaPS) project funded by the Norwegian Programme for Development, Research and Education (NUFU), NUFUPRO-2007/10121. The authors thank the National Meteorological Agency of Ethiopia for allowing the first author to carry out this research and for providing Ethiopia rainfall data.