Analysis of the Performance of the 2017 Agricultural Season in Mali

Analysis of the Performance of the 2017 Agricultural Season in Mali March 2018 www.africanriskcapacity.org

Contents Contents... 2 1. Context... 3 2. Performance of the 2017 Agricultural Season... 4 2.1. Preliminary evaluation of the 2017 agricultural season by the SAP... 4 2.2. Cadre Harmonisé... 5 2.3. Africa RiskView Estimates... 6 2.4. Overview of Potential Discrepancies... 7 3. Rainfall... 8 3.1. Methodology and Limitations... 8 3.2. Overall Observations... 8 3.2.1. Cumulative Rainfall... 8 3.2.2. Temporal Distribution... 9 3.2.3. Deficit Areas... 11 4. Conclusions... 16 4.1. Planting Dates... 16 4.1.1. Methodology and Limitations... 17 4.1.2. Overall Observations... 17 4.1.3. Conclusions... 22 4.2. Crop Cycle Length... 22 4.2.1. Methodology and Limitations... 22 4.2.2. Overall Observations... 22 4.2.3. Conclusions... 23 4.3. Benchmark... 24 4.3.1. Methodology and Limitations... 24 4.3.2. Overall Observations... 24 4.3.3. Preliminary Conclusions... 24 4.4. Summary of Conclusions... 24 4.4.1. Rainfall... 24 4.4.2. Planting Windows... 24 4.4.3. Crop Cycle Length... 25 4.4.4. Benchmark... 25 4.4.5. Combination of explanations... 25 Annex I... 26 Page 2

1. Context In October 2017, the Government of Mali expressed concerns over the performance of the 2017 agricultural season. This interpretation of the progression of the season was confirmed by the Mali Technical Working Group (TWG). At the same time, concerns were raised over the ability of Africa RiskView to accurately pick up the number of people affected, as modelled estimates for the 2017 season were around 400,000 people, whereas initial information suggested that the number of food insecure people in 2017/18 might be higher. A ground-truthing mission was conducted in early December 2017, to assess the situation on the ground comparing government data and data provided by partners, and discuss potential discrepancies between Africa RiskView s estimates and the information collected by Government institutions. The ground-truthing mission resulted in the following observations regarding the performance of the 2017 agricultural season, and the ability of Africa RiskView to accurately pick up the developments on the ground: According to the information presented by the TWG, the 2017 agricultural season was marked by a poor distribution of rainfall in some regions. Specifically, a false start of the season in May 2017 and a dry-spell in September 2017 were highlighted as having affected crucial stages of crop development. Rainfall data from 42 stations across Mali was provided to the ARC team. Due to the poor performance of the 2017 rainy season in some regions, the performance of the agricultural season was expected to be worse than the official production figures issued by the Ministry of Agriculture, which were based on projections that did not take into account the effect of the September dry-spell and suggested that agricultural production had increased by over 5% compared to 2016, and by over 30% compared to the 5-year average. Instead, the Mali TWG suggested to use a map with preliminary results of the agricultural season, which indicated that large parts of Kayes and Mopti regions, as well as parts of Koulikoro and Gao regions were affected by a poor or very poor season. The official results from the Cadre Harmonisé consensus-based food security classification exercise, which includes the Government, UN agencies and NGOs, indicate that around 800,000 people are likely to be food insecure during the peak lean season. Informal discussions with the Mali TWG and partners suggested that these projections were too optimistic, and that the actual number of food insecure people would likely be corrected upwards during the review of the Cadre Harmonisé figures in March 2018. Based on the findings above, the mission concluded that the potential discrepancies between the modelled estimates and the situation on the ground should be investigated further, to assess whether a recustomisation of Africa RiskView would be required. Based on three key parameters, an analysis plan was agreed upon with the Government, to address the following: 1) A comparison of satellite estimates and station data to assess potential discrepancies between the two, based on the station data for 42 stations provided by Mali Météo. 2) A comparison of modelled and actual planting dates for the 2017 season, given that initial observations from the ground seemed to indicate that there was a discrepancy between the planting windows in Africa RiskView and the actual planting dates. 3) A comparison of crop varieties, and specifically their cycle length, or Length of Growing Period (LGP), given the paramount importance of this parameter in the crop modelling. 4) A review of the benchmark used in Africa RiskView, given that the one-year benchmark used currently might not be the most appropriate benchmark. A second mission took place between 13 and 15 February 2018 to present the preliminary findings of this analysis to, and discuss the way forward with, the TWG, and to gather additional data in order to allow a more refined analysis (particularly in terms of sowing dates). This data collection was achieved through the Page 3

submission of specific questionnaires to the Regional Directors for Agriculture who were requested to provide answers at sub-regional level. The following report summarises the findings of this analysis, which was conducted between December 2017 and March 2018, and develops some recommendations for the consideration of ARC Agency s and ARC Ltd s Senior Management. 2. Performance of the 2017 Agricultural Season 2.1. Preliminary evaluation of the 2017 agricultural season by the SAP According to the information provided to ARC by the Mali TWG, the performance of the 2017 agricultural season was average overall, with a poor to very poor seasonal performance in some areas. The rainy season was characterised by a normal to early start throughout the country, followed by poor and erratic rainfall in June and early July 2017. While normal rainfall was recorded between mid-july and early September, the season ended earlier than normal in mid-september, with very little rainfall in October 2017. Based on these observations, the rainfall was, according to the TWG, below normal, and poorly distributed in time and space, resulting in a below average seasonal performance in some areas. The following map, produced by the Early Warning System (Système d Alerte Précoce, SAP), highlights areas that were affected by a below normal season; the most affected areas include the northern parts of Kayes region, central Koulikoro and western Mopti: Figure 1: Preliminary evaluation of the 2017 agricultural season (source: SAP) Despite the existence of additional causes of food insecurity (mentioned below), the map above was subsequently confirmed to be the best available representation of the impact of drought on rainfed agriculture. Based on this preliminary assessment of the agricultural season, the SAP identified areas at risk of food insecurity during the 2017/18 lean season. These include 25 communes at high risk of food insecurity (difficultés alimentaires) due to a combination of poor agro-pastoral production, low river levels in the delta, and the impact of civil insecurity; the identified high-risk areas are located in central Mopti and northern Kayes. Page 4

An additional 72 communes were identified as being affected by severe economic difficulties (difficultés économiques sévères); these are located predominantly in northern Kayes, central Koulikoro and western Mopti, and also affected by poor agro-pastoral production and the impact of civil insecurity. Finally, 204 communes were identified as being affected by mild economic difficulties (difficultés économiques légères); these are areas affected by a decrease in agro-pastoral production, high food prices and the impact of civil insecurity, and are located mainly in Kayes, parts of Koulikoro, Mopti, Tombouctou, Kidal and Gao, as the following map illustrates: Figure 2: Food insecurity risk during the 2017/18 season (source: SAP) In terms of affected populations, no information on the number of food insecure and/or drought affected people as estimated by the SAP was available at the time of this analysis. 2.2. Cadre Harmonisé The Cadre Harmonisé is a consensus food security classification exercise based on the Integrated Food Security Phase Classification (IPC) methodology, which includes all relevant stakeholders (incl. Government, UN agencies, NGOs etc.) and aims at providing an evidence-based projection of the food security situation in the country after each season. The final figures for the 2017/18 season indicate that nearly 800,000 people are likely to be food insecure (phases 3-5 1 ) at the peak lean season in June-August 2018, as a result of the performance of the 2017 agricultural season, and other compounding effects (such as civil insecurity etc.). The cercles most affected by food insecurity are Kolokani in Kayes region with 18% of the population in phases 3-5, Tenenkou (21%), Dienné (14%) and Mopti (12%) in Mopti region, as well as Ménaka (12%) in Gao region. 1 The assumption can be made that the classification in phases 3-5 roughly corresponds to the two most severe categories in the SAP s classification exercise (high risk of food insecurity and severe economic difficulties), while phase 2 (stressed) should correspond to the mild economic difficulties category. Page 5

Figure 3: Projected IPC classification for June-August 2018 (source: Cadre Harmonisé) It is important to note that, according to the TWG and external partners consulted during the ground-truthing mission in Bamako in December 2017 indicated, these projections are too optimistic, and likely to be revised upwards during the revision of the Cadre Harmonisé analysis in March 2018. 2.3. Africa RiskView Estimates At the end of the 2017 agricultural season, Africa RiskView estimates that around 400,000 people were affected by drought in Mali. These are located mainly in northern and eastern Kayes region (Nioro, Kita, Diéma, Bafoulabé and Kayes); northern Koulikoro (Nara and Kolokani); parts of Segou (Macina and Niono); as well as Ansongo in Gao region: Figure 4: Estimated population affected by drought after the 2017 season (source: Africa RiskView) Page 6

It appears that Africa RiskView picks up some areas affected by a poor season according to the SAP and the Cadre Harmonisé (e.g. parts of Kayes), but does not trigger a drought event in other areas, such as parts of Mopti, which were severely affected as per the SAP s analysis and also show a large population in food insecurity according to the Cadre Harmonisé. Vice versa, some areas identified by Africa RiskView, such as Kita or Nara, seem to be only mildly affected by a poor agricultural season and food insecurity according to the SAP and the Cadre Harmonisé. The ground-truthing process and review of the Africa RiskView customisation focused on analysing these potential discrepancies between modelled estimates and the situation on the ground, based on the analysis plan developed with the Mali TWG in December 2017. 2.4. Overview of Potential Discrepancies The following table provides an overview of the classifications and population affected estimates for each source of information available at the time of this analysis, and highlights potential discrepancies between these sources. In comparing these figures, it is however crucial to keep in mind the methodological differences between the three approaches. Region Cercle Performance of 2017 Agricultural Season (Mali SAP) Cadre Harmonisé: CH Phase Pop. in Phase 3-5 Africa RiskView Estimate Potential Discrepancy Bafoulabé Poor/Very Poor Stressed 9,106 6,380 Diéma Poor Stressed 27,512 48,512 Kayes Poor/Average Minimal 13,334 - Discrepancy Kayes Keniéba Average/Good Minimal - - Kita Average Minimal 11,238 48,433 Nioro Poor/Very Poor Stressed 20,819 94,518 Yélimané Poor Minimal 4,586 9,706 Banamba Poor/Average Stressed 22,335 - Discrepancy Dioïla Average/Good Minimal - - Kangaba Average/Good Minimal 1,304 - Koulikoro Kati Poor/Very Poor Stressed 62,155 - Discrepancy Kolokani Poor/Very Poor Stressed 54,375 8,972 Koulikoro Poor/Average Minimal 2,736 - Discrepancy Nara Average Minimal 9,429 94,796 Bamako Bamako Poor Minimal 105,551 - Discrepancy Barouéli Good Minimal 7,907 - Bla Good Minimal - - Macina Average Minimal 6,134 82,190 Segou Niono Average Minimal 9,481 5,750 San Average Minimal 4,334 - Ségou Average Minimal 18,088 - Tominian Average Stressed 17,238 - Bougouni Average Minimal - - Kadiolo Average Minimal 6,325 - Kolondiéba Average Minimal - - Sikasso Koutiala Average Minimal - - Sikasso Average Minimal 9,549 - Yanfolila Average Minimal - - Yorosso Average Minimal - - Bandiagara Poor/Average Stressed 36,646 - Discrepancy Bankass Average/Good Minimal 3,439 - Dienné Poor Stressed 37,902 - Discrepancy Mopti Douentza Average Stressed 25,629 - Koro Average/Good Minimal 23,550 - Mopti Poor Stressed 57,504 - Discrepancy Tenenkou Poor Crisis 44,443 - Discrepancy Youwarou Poor Stressed 9,868 - Discrepancy Diré Average Stressed 5,701 - Goundam Poor Stressed 19,668 - Discrepancy Tombouctou Gourma-Rharous Average Stressed 10,101 - Niafunké Average Stressed 13,681 - Tombouctou Average Stressed 18,203 - Ansongo Poor Stressed 13,723 7,197 Gao Bourem Average Stressed 10,589 - Gao Average Stressed 28,026 - Ménaka Average Stressed 8,495 - Page 7

As can be seen, these discrepancies concern mostly Koulikoro and Mopti. 3. Rainfall The first step of the analysis plan included the comparison of satellite estimates and station data, to assess potential discrepancies between the two datasets, using rainfall station data for 42 stations provided by Mali Météo. The comparative analysis focused on determining whether the rainfall measured on the ground (both in terms of cumulative seasonal rainfall and the spatial and temporal distribution of the rains) was accurately picked up by the satellite rainfall dataset chosen by the Mali TWG during the customisation of Africa RiskView. 3.1. Methodology and Limitations The dekadal rainfall station data for 42 stations provided by Mali Météo was imported into Africa RiskView, and the corresponding pixel-level satellite rainfall estimates for all available dataset exported from the software. The different datasets (stations and satellite) were then compared, focusing both on cumulative seasonal rainfall for the period May-October 2017 and the temporal distribution of the rains over the season. This analysis was conducted at the regional level (averaging all stations/corresponding pixels in a given region), given that satellite rainfall estimates tend to be more accurate at the aggregated level, rather than at the pixel level. It is important to note that due to the inherent differences in methodology, no perfect match can or should be expected between station and satellite data. Indeed, the station data provides an actual measurement of rainfall, while the satellite data estimates the rainfall based on remotely sensed information like cloud temperature. More importantly, the granularity of the data also needs to be taken into consideration, given that the pixel-level estimates are made for an area of roughly 10 km by 10 km, while the station data only measures rainfall in a given location, whichmay not always reliably be extrapolated to a larger area, such as the one for which the satellite estimates apply. In other words, the station data may provide a much more accurate measurement of actual rainfall in a given point, but the extent to which that measurement is representative of actual rainfall in the area around that point is another question and arguably the satellite data may provide a better estimate of average rainfall than station-based interpolations. Because of these inherent differences, this point-to-pixel comparison cannot result in any definite conclusion about the performance of the satellites. Instead, it can only suggest where further investigation might be needed. 3.2. Overall Observations 3.2.1. Cumulative Rainfall In terms of cumulative rainfall, the 2017 seasonal rainfall measured on the ground appears to be well picked up by the satellite dataset selected by the Mali TWG for the customisation of Africa RiskView (ARC2), with a good match for 4 out of 6 regions. However, ARC2 overestimates the station rainfall in Kayes and Mopti regions, by 17% and 31%, respectively. Compared to other satellite datasets, ARC2 performs well; CHIRPS seems to perform better in terms of cumulative rainfall, but also overestimates the rain in Kayes region (by 11%). The following table shows the cumulative rainfall, and comparison to station data, for all tested datasets: CUMULATIVE RAINFALL ANOMALY STATION ARC2 RFE2 CHIRP CHIRPS TAMSAT ARC2 RFE2 CHIRP CHIRPS TAMSAT Bamako 796 771 723 772 846 699-3% -9% -3% 6% -12% Kayes 596 695 737 561 663 548 17% 24% -6% 11% -8% Koulikoro 764 758 745 705 775 620-1% -2% -8% 2% -19% Mopti 497 651 676 422 517 421 31% 36% -15% 4% -15% Segou 670 659 688 521 617 488-2% 3% -22% -8% -27% Sikasso 904 876 1,004 826 862 697-3% 11% -9% -5% -23% Page 8

With these discrepancies in mind, the same analysis has been conducted for 2016 to find out if these overestimations are consistent over time, and to potentially allow correction of benchmark estimates in an adequate manner. The table below shows average differences between cumulative rainfall data as per station data compared with the available satellite datasets. It shows that there are significant differences between 2017 and 2016, and any correction on the rainfall data would have to be applied year by year. CUMULATIVE RAINFALL ANOMALY STATION ARC2 RFE2 CHIRP CHIRPS TAMSAT ARC2 RFE2 CHIRP CHIRPS TAMSAT Bamako 911 1,044 1,139 879 895 733 15% 25% -3% -2% -20% Kayes 762 813 869 689 750 656 7% 14% -10% -2% -14% Koulikoro 812 901 974 782 806 655 11% 20% -4% -1% -19% Mopti 632 650 755 467 564 433 3% 19% -26% -11% -31% Segou 717 733 724 583 645 545 2% 1% -19% -10% -24% Sikasso 1,104 1,062 1,045 899 978 696-4% -5% -19% -11% -37% 3.2.2. Temporal Distribution The visual analysis of the distribution of the rains at the regional level confirms the pattern above, with 4 out of 6 regions showing a close match between the station and satellite data: the overall trends are picked up well for Bamako, Koulikoro, Segou and Sikasso. In line with the observations about the cumulative rainfall above, there is a significant overestimation of rainfall for Kayes and Mopti; in both cases, this overestimation is relatively consistent throughout the season: Page 9

Page 10

These charts show that the misestimation by the satellite sensor is quite variable throughout the season (only Kayes shows a relatively constant overestimation from dekad to dekad) and between regions (the timing of the highest overestimations and underestimations differ from region to region). Therefore, it is not clear whether what would be achieved by applying a systematic rainfall correction factor (through the Percentage of Effective Rainfall parameter in Africa RiskView) would actually improve estimates of drought severity, since the correction factor would have to affect all rainfall estimates equally. (For example in the case of a reduction factor to compensate for an average overestimation, underestimations would be exacerbated.) 3.2.3. Deficit Areas Based on the findings of the regional-level analysis, the match between satellite and station data was analysed for specific stations located in deficit areas as identified by the SAP in Mopti and Kayes and in some other areas, to understand whether an overestimation of rainfall might explain potential differences between the model and the situation on the ground. Overall, 13 out of the 42 stations are located in areas identified by the SAP as having experienced a poor or very poor agricultural season in 2017. This includes 6 stations in Kayes (Bafoulabe, Diema, Kayes, Mahina, Nioro du Sahel and Yelimane), 4 stations in Koulikoro (Didieni, Kati-Haut, Kolokani and Ouelesseb), 1 station in Mopti (Mopti), 1 station in Segou (Tominian) and 1 station in Sikasso (Yorosso). Similar to the regional analysis, the comparison of satellite estimates and station data for these 13 stations shows a mixed picture. In Kayes region, ARC2 data appears to systematically overestimate the rainfall measured on the ground, with the exception of a slight underestimation of station rainfall in late August and early September for the stations in Diema and Kayes: Page 11

Page 12

In Koulikoro region, the satellite estimates show a relatively good match in one stationand an underestimation of the rainfall in two stations, but in Kolokani, which according to the SAP is an area that was affected by a very poor agricultural season, ARC2 significantly overestimates the station data, particularly between May and August 2017; this might explain why Africa RiskView did not pick up a significant drought in this area. Page 13

Page 14

In Mopti, one station is located in a deficit area based on the SAP s information. It appears that the satellite data overestimates the rainfall for this station, particularly during certain dekads in June and July 2017: For Tominian station in Segou region, there is a relatively good match between the station and the satellite data: Finally, for Yorosso station in Sikasso region, the satellite data also matches the station data quite well, as the following graphs illustrate: Page 15

4. Conclusions The comparison of station and satellite data for Mali shows that overall, in 4 out of 6 regions, there is a relatively good match between the satellite and the station data; the cumulative rainfall, as well as the spatial and temporal distribution of the rains appears to be well picked up. However, the satellite estimates generally overestimate the rainfall measured in Kayes and Mopti regions, as well as for one station located in Koulikoro region. According to the SAP, most of these areas were affected by a poor or very poor agricultural season in 2017, which was only partly picked up by Africa RiskView (specifically in Diema, Nioro du Sahel and Yelimane), which suggests that satellite overestimation of rainfall might in part explain why Africa RiskView didn t detect a drought in some regions. However, drawing definite conclusions on the actual mismatch between satellite data and actual rainfall is complicated by the fact that point-to-pixel comparisons are inherently flawed (because satellite data refer to an average value over a very large area while station data refer to a very small area), as well as the fact that these mismatches are not consistent over time or space. Correcting for any misestimation in Africa RiskView would only be possible by reducing the effective rainfall percentage for each area depending on the average overestimation observed. This approach would assume that this misestimation affects the entire region in equal manner, and it would affect all satellite data, reducing overestimations and exacerbating underestimations or vice versa depending on the case. If this were to be done, it would be relevant to do this consistently: for all areas (regardless of the amount of misestimation) and applying this approach to previous years as well to make sure that the benchmark is modelled in a comparable way. The result of such correction would be a significant increase in estimated numbers of affected people in Kayes, a reduction in the number in Koulikoro, and a very big increase in Mopti (from 0 to approx. 900,000 people affected). Comparing this with the SAP information indicates that the severity of the impact seems appropriate for Kayes, underestimated in Koulikoro, and overestimated in Mopti and Segou. This will be further discussed in the last section. 4.1. Planting Dates The second step of the analysis plan developed with the Mali TWG in December 2017 focused on the comparison of actual planting dates for 2017 and the planting windows defined during the customisation of Africa RiskView. Information on the actual planting dates was compared to the modelled plantings for 2017, to identify discrepancies and assess the potential impact of these discrepancies on the modelled estimates. Page 16

4.1.1. Methodology and Limitations Information on actual planting dates disaggregated by Cercle was collected from the Regional Agriculture Directors through an ad hoc questionnaire and follow-up phone interviews. Those dates have been compared to the planting windows in Africa RiskView, and to the modelled planting dates for the 2017 season, as well as with other sources of information (only available at regional level). The comparison of actual and modelled planting dates has a number of limitations which are important to note. First, information on actual planting dates might not be entirely accurate, particularly given that some level of detail might be lost in the aggregation of information at the cercle level. Moreover, actual plantings can be affected by a variety of factors (incl. access to agricultural inputs etc.) that would not be taken into account in Africa RiskView, which uses only pre-defined rainfall thresholds to determine planting dates. 4.1.2. Overall Observations The planting windows defined during the customisation of Africa RiskView for Pool 2 and subsequent risk pools vary by region/agro-ecological zone. In the southern parts of the country, the planting window starts in May (dekad 15) and ends in early July (dekad 19); in the central parts of Mali, the window starts in early June (dekad 16) and lasts until mid-july (dekad 20); and in the northern agricultural areas, it lasts from mid-june (dekad 17) to the end of July (dekad 21). Start of Planting Window End of Planting Window Initial discussions during the ground-truthing mission in December 2017 suggested that, in 2017, the actual planting dates fell outside these pre-defined planting windows; the timing of the planting was influenced by a variety of factors, including the distribution of rainfall (and particularly the false start of the season in early May), access to agricultural inputs (such as seeds and tractors), and other external factors. To assess the potential impact of reflecting the actual planting dates in Africa RiskView, sowing dates by cercle were collected along with other information (see annex 1), which suggests that planting dates across the different regions varied significantly: In Kayes region, planting happened mostly toward the end of the sowing window, and in some cases (like in Kayes, Kita and Yelimane) after the sowing window; In Koulikoro region, planting was reported to have spanned across 6 to 7 dekads, including 1 to 3 dekads after the end of the sowing window (depending on the cercle); Page 17

In Bamako region, planting allegedly occurred only very early, at the beginning of May (which is 2 dekads before the beginning of the normal sowing window); In Segou region, planting dates varied significantly from one Cercle to another, starting as early as mid- May (in Bla) and going into beginning July (in Niono and Segou); In Sikasso region, planting happened almost throughout the entire normal sowing window, but extended 1 to 3 dekads after the end of the window in most Cercles; In Mopti region, the situation was also quite varied, with some Cercles planting early (beginning and mid- June, while others planted late (from mid-july to mid-august), and other throughout the season; InTombouctou region, planting occurred in the middle of the sowing window (between late June and early July). No data was obtained for Gao region so the estimated sowing dates were not changed from Africa RiskView s best guess. The comparison of these actual planting dates with the planting windows in Africa RiskView and the modelled planting dates for the 2017 season highlight some significant discrepancies. Overall, it appears that actual plantings in 2017 took place in a much wider timeframe than the planting windows determined during the customisation, which extend over 5 dekads throughout the country. Early plantings have been reported in various Cercles although this was reported as an early planting in only one Cercle, suggesting that the normal sowing window might need to be reviewed to start earlier in some areas. Similarly, several late plantings have been reported, but not always labelled as such, meaning again that the sowing window in the Mali customisation of the Drought Model might need to be reviewed to include these dates as part of normal practice unless the circumstances leading to this practice in 2017 were relatively exceptional and unlikely to be repeated in future years. The following table shows the difference between reported planting dates and Page 18

1-10 mai 13 11-20 mai 14 21-31 mai 15 1-10 juin 16 11-20 juin 17 21-30 juin 18 1-10 juillet 19 11-20 juillet 20 21-31 juillet 21 1-10 Aout 22 11-20 Aout 23 21-31 Aout 24 REGION CERCLE Avant mai (date) the sowing windows discussed by the TWG and used in the customisation of the Drought Model. Reported Sowing Dates + comparison with the sowing window in ARV (dates that fall outside the ARV sowing window are in yellow) Mai 2017 Juin 2017 Juillet 2017 Août 2017 Bafoulabé Diéma Kayes Kayes Keniéba Kita Nioro Yélimané Banamba Dioïla Kangaba Koulikoro Kati Kolokani Koulikoro Nara Bamako Bamako Barouéli Bla Macina Segou Niono San Ségou Tominian Bougouni Kadiolo Kolondiéba Sikasso Koutiala Sikasso Yanfolila Yorosso Bandiagara Bankass Djenné Mopti Douentza Koro Mopti Tenenkou Youwarou Diré Goundam Tombouctou Gourma-Rharous Niafunké Tombouctou Ansongo Gao Bourem Gao Ménaka Page 19

This significant discrepancy between the planting windows in Africa RiskView and the actual planting dates in 2017 could be addressed by changing the planting windows and parameters 2 in the model. Until the relevance of expanding the normal sowing windows accordingly is confirmed, it will be assumed that the current sowing windows are still valid for other years, and in particular for 2016 (the benchmark reference year). Another analysis was conducted to compare reported sowing dates with the estimated sowing dates in Africa RiskView. This is presented in the table below. 2 Depending on whether specific dates or a new window have been reported, the sowing criteria used in Africa RiskView to determine if the conditions are right for sowing may have to be removed: if sowing is reported to have taken place in a given dekad, it is assumed that sowing took place, but if instead a whole sowing window has been reported (no specific dates), then it is assumed that sowing actually took place during that window only when rainfall is sufficient, so the sowing criteria still need to be applied. Page 20

1-10 mai 13 11-20 mai 14 21-31 mai 15 1-10 juin 16 11-20 juin 17 21-30 juin 18 1-10 juillet 19 11-20 juillet 20 21-31 juillet 21 1-10 Aout 22 11-20 Aout 23 21-31 Aout 24 REGION CERCLE Avant mai (date) Favourable sowing conditions in ARV (in % of Cercle area) + comparison with reported sowing dates (reported sowings in green and in orange; unused favourable sowing opportunites in yellow) Mai 2017 Juin 2017 Juillet 2017 Août 2017 Bafoulabé 13% 7% 98% 29% 31% 100% 68% 100% 99% 95% 99% Diéma 0% 0% 69% 11% 5% 100% 100% 100% 85% 96% 30% Kayes 0% 0% 27% 18% 15% 92% 29% 100% 91% 99% 100% Kayes Keniéba 2% 59% 98% 37% 66% 100% 98% 100% 97% 100% 100% Kita 1% 40% 99% 46% 78% 100% 98% 100% 100% 58% 99% Nioro 0% 0% 38% 1% 12% 97% 100% 100% 67% 99% 22% Yélimané 0% 0% 16% 4% 16% 98% 90% 100% 92% 100% 98% Banamba 0% 0% 93% 15% 35% 35% 100% 100% 100% 100% 97% Dioïla 75% 85% 100% 61% 100% 58% 100% 100% 100% 100% 100% Kangaba 21% 58% 100% 97% 100% 100% 100% 100% 100% 100% 87% Koulikoro Kati 9% 62% 99% 86% 92% 90% 100% 100% 100% 93% 100% Kolokani 0% 17% 96% 47% 53% 90% 100% 100% 100% 97% 98% Koulikoro 6% 17% 100% 60% 79% 11% 100% 100% 100% 100% 100% Nara 3% 2% 25% 0% 5% 72% 98% 100% 74% 96% 16% Bamako Bamako 0% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% Barouéli 77% 44% 97% 3% 72% 33% 100% 100% 100% 100% 100% Bla 60% 16% 90% 14% 40% 6% 100% 90% 100% 100% 100% Macina 22% 0% 59% 0% 27% 0% 100% 100% 86% 86% 14% Segou Niono 23% 0% 20% 0% 53% 4% 100% 100% 92% 92% 25% San 22% 5% 95% 18% 45% 16% 100% 100% 100% 100% 78% Ségou 34% 0% 61% 0% 30% 27% 97% 100% 100% 100% 86% Tominian 21% 0% 75% 9% 95% 16% 100% 100% 100% 100% 43% Bougouni 91% 85% 100% 55% 100% 87% 99% 100% 100% 100% 99% Kadiolo 100% 100% 100% 96% 100% 100% 100% 100% 100% 100% 100% Kolondiéba 94% 89% 100% 85% 100% 97% 100% 100% 100% 100% 100% Sikasso Koutiala 81% 11% 96% 91% 75% 37% 100% 100% 100% 100% 97% Sikasso 80% 91% 97% 99% 100% 84% 100% 100% 100% 100% 97% Yanfolila 97% 75% 100% 89% 96% 100% 99% 100% 100% 93% 99% Yorosso 26% 0% 41% 95% 92% 100% 100% 100% 100% 100% 74% Bandiagara 2% 17% 0% 0% 28% 0% 100% 100% 100% 100% 20% Bankass 2% 59% 0% 6% 83% 33% 100% 100% 100% 100% 13% Djenné 0% 0% 84% 0% 30% 5% 100% 100% 92% 92% 8% Mopti Douentza 4% 17% 5% 0% 37% 0% 100% 100% 98% 100% 72% Koro 0% 62% 13% 5% 72% 29% 95% 100% 100% 100% 63% Mopti 10% 0% 3% 0% 31% 7% 98% 100% 100% 100% 20% Tenenkou 12% 0% 34% 0% 29% 0% 100% 100% 97% 97% 18% Youwarou 1% 0% 0% 0% 73% 18% 100% 100% 99% 100% 85% Diré 0% 0% 0% 0% 56% 0% 100% 100% 100% 100% 94% Goundam 2% 0% 0% 0% 55% 0% 75% 81% 91% 100% 44% Tombouctou Gourma-Rharous 1% 0% 0% 0% 25% 0% 98% 100% 100% 100% 38% Niafunké 0% 0% 0% 0% 74% 12% 100% 100% 100% 100% 76% Tombouctou 1% 0% 0% 0% 32% 0% 93% 34% 75% 99% 3% Ansongo 0% 0% 0% 1% 80% 0% 94% 43% 100% 100% 63% Gao Bourem 0% 1% 0% 0% 1% 10% 85% 70% 98% 88% 23% Gao 0% 15% 2% 1% 32% 5% 98% 95% 97% 52% 28% Ménaka 0% 0% 3% 4% 93% 0% 63% 13% 99% 98% 63% Page 21

In most cases the reported sowing dates correspond to a valid sowing date (in green). However, several reported sowing dates (especially some first sowings) seem to have occurred when rainfall was not conducive to planting (in orange). Vice versa, there are a lot of potential sowing opportunities (where minimum rainfall criteria have been met according to Africa RiskView and satellite data) that have not been taken advantage of by the farmers (in yellow). This shows or confirms that farmers behaviour (and specifically the decision to plant) is driven by other factors than just rainfall, and the importance of the rainfall criteria will vary. Some reported sowing dates (like the one for Bamako) seem particularly unlikely when compared to both the sowing window, the rainfall and the regional-level data on sowing dates, but considering the complexity of farmers behaviour the reported sowing dates have net been altered. 4.1.3. Conclusions The comparison of reported and modelled planting dates highlights a significant discrepancy between Africa RiskView and the situation on the ground. In some cases, actual plantings occurred outside of the planting windows defined by the Mali TWG during the customisation of the model. Even though there appears to be some discrepancies and/or incoherencies in the data obtained, it would be quite risky to assume that straightforward corrections could be made to the data in an attempt to provide a more realistic picture of actual plantings. Therefore, if this discrepancy is addressed by using the actual planting dates in the model, this would lead to a small increase in Africa RiskView s estimates for Kayes and Mopti, and a very big increase in Koulikoro (from approx. 100,000 to 400,000 people affected) and in Bamako (from 0 to 240,000). Comparing this with the SAP information indicates that the severity of the impact seems appropriate for Kayes and Mopti (although not located in the right Cercles), and overestimated in Koulikoro, Bamako and Sikasso. These differences seem to indicate that sowing dates alone cannot explain all the differences between the field and the model. 4.2. Crop Cycle Length The third step of the analysis plan consisted of a comparison of the cycle lengths of the maize varieties planted by farmers in 2017, and the Length of Growing Period (LGP) defined in Africa RiskView. The information on maize varieties was also obtained from the Regional Agriculture Directors through the survey. 4.2.1. Methodology and Limitations The information on cycle lengths was compared to the LGPs determined during the customisation of Africa RiskView. The analysis was limited mainly by the fact that in most Cercles only the most prevalent LGP was reported, or in some cases a range was provided but no information about the prevalence of each. No information has been obtained for Sikasso and Gao therefore the values currenly used in the customisation (respectively 120 days and 90 days) have been maintained. Since the LGP is not something that is likely to change significantly from year to year, this analysis would require using the LGP data both for the current year and for the benchmark. 4.2.2. Overall Observations Data from Mali Météo and the Mali Agro-Meteorological Service was used to determine the LGP during the customisation of Africa RiskView for Pool 2. This data suggests that in southern Mali, a long cycle maize crop (120 days) is prevalent, while the LGP is shorter in the central and northern parts of the country (90 days). The LGP was aligned with the planting windows, meaning that areas where the season starts earlier have a longer LGP and vice versa for areas where the season starts later. During the December 2017 ground-truthing mission it was suggested that, nowadays, farmers may be using a lot of short-cycle varieties (70-75 days to maturation) in some regions, but data obtained through the survey revealed that farmers reportedly use shorter-cycle crops in the south (104 days on average) but do indeed use crops with an average LGP of 92 days in the north. In half the Cercles were it was thought that 120-day maize was being used, it was reported that farmers use Page 22

Reported LGP Customisation LGP Reported LGP REGION CERCLE REGION CERCLE Customisation LGP mostly varieties of 100 days or less. These differences between what is currently being used in the model and what has been reported through the survey suggests that it might be necessary to review this parameter during the next customisation exercise. Cycle length (differences of >20 days in yellow) Cycle length (differences of >20 days in yellow) Bafoulabé 120 120 Diéma 100 90 Kayes 90 90 Kayes Keniéba 90 120 Kita 110 120 Nioro 90 90 Yélimané 90 90 Banamba 80 90 Dioïla 110 120 Kangaba 110 120 Koulikoro Kati 110 120 Kolokani 80 90 Koulikoro 80 120 Nara 80 90 Bamako Bamako 115 120 Barouéli 100 120 Bla 100 120 Macina 100 90 Segou Niono 100 90 San 100 120 Ségou 100 90 Tominian 100 120 Sikasso Mopti Tombouctou Gao Bougouni 120 Kadiolo 120 Kolondiéba 120 Koutiala 120 Sikasso 120 Yanfolila 120 Yorosso 120 Bandiagara 95 90 Bankass 90 90 Djenné 90 90 Douentza 120 90 Koro 85 90 Mopti 90 Tenenkou 70 90 Youwarou 90 90 Diré 95 90 Goundam 95 90 Gourma-Rharous 100 90 Niafunké 100 90 Tombouctou 95 90 Ansongo 90 Bourem 90 Gao 90 Ménaka 90 The impact of such change in the model will vary depending on the prevailing rainfall patterns: shorter cycle varieties will be more sensitive to dry spells while longer cycle varieties will be more at risk of suffering from an early end of the rains. Therefore, the actual sowing dates will also be critical in determining how each of these is affected by drought. 4.2.3. Conclusions While the question of the LGP refers mainly to a question of customisation choice, it is still an important one to assess as there are specific interactions with the sowing dates in particular. The survey was also used to try to find out if there have been recent trends in the choice of the LGP, and the responses suggest that this might indeed be the case in Koulikoro. However, considering that the benchmarks is simply the past year, it is unlikely that this trend would have any impact on the model s estimates. Although the difference might be more significant in other years (i.e. under different rainfall patterns), in the case of 2017, switching the LGP to the reported values only has a significant impact on Segou estimates which increase by approx. 150,000 people. In comparison with the SAP data, this is actually less in line with what is expected for Segou. This doesn t mean that the reported LGP values are not correct, but instead that other changes would have to be made in parallel. Page 23

4.3. Benchmark The final step of the analysis plan consisted in reviewing the benchmark used as an indicator for normal conditions in Mali. During the customisation of Africa RiskView for Pool 3, the TWG decided to a one-year benchmark, which means that the previous year is used as reference to calculate the drought severity at the end of each agricultural season in the model. While this short benchmark allows to pick up more recent drought events, it bears the risk of triggering false positives or negatives (i.e. triggering drought events that did not occur in reality, and vice versa), particularly in cases where a season with a normal agricultural performance on the ground is preceded by a very good or bad year. Notwithstanding the decision by the Mali TWG to choose a short benchmark in order to de-trend the response costs series, using a longer benchmark would arguably produce more accurate estimates of people affected. 4.3.1. Methodology and Limitations As part of the review of Africa RiskView, a number of WRSI parameter combinations were tested to assess the impact of each parameter change on the modelled estimates, as discussed above. For each newly created project, the drought severity was determined using the 2016 season as benchmark in the context of the benchmark analysis, this calculation was reproduced using a 5-year benchmark instead. The main limitation of this analysis is that the traditional model performance indicators (such as the correlation with yield or historical data) are not taken into consideration. A more in-depth review of the benchmark should take these, as well as the historical MDRC series, into account. A change in benchmark would also require a revision of the risk transfer parameters. 4.3.2. Overall Observations Given the good performance of the 2016 season, the switch from a 1-year benchmark to a 5-year benchmark results in a decrease of the population affected estimates in Africa RiskView. On average, for all projects, the numbers of affected people decrease by around 100-200,000 people. 4.3.3. Preliminary Conclusions Based on the observations above, a switch from a 1-year to a 5-year benchmark would result in a decrease of the numbers of people affected. This decrease is consistent at sub-national level, which means that the modelled estimates decrease in all regions. 4.4. Summary of Conclusions 4.4.1. Rainfall The comparison of satellite estimates and data from 42 rainfall stations across Mali for the 2017 agricultural season suggests that, in most areas, the rainfall measured on the ground was accurately picked up by the satellite dataset chosen by the Mali TWG for the customisation of Africa RiskView (ARC2). However, it does appear that the ARC2 data overestimates the rainfall measured in Kayes and Mopti regions, as well as parts of Koulikoro. Inherent limitations in the point-topixel comparison prevent any definite conclusion, but it seems likely that there was an actual overestimation of rainfall in the satellite data, and this in turn would have contributed to Africa RiskView s underestimation of the impact of the 2017 drought. The relevance of correcting this overestimation by reducing the effective rainfall percentage is a subject of discussion, but such a correction would clearly result in an increase in the modelled estimates. 4.4.2. Planting Windows It appears that the actual planting dates in 2017 differ significantly from the planting windows defined during the customisation of Africa RiskView, and the modelled planting dates. In some instances (e.g. in Koulikoro and Sikasso), actual plantings occurred far outside the planting windows in the model, and often didn t occur as soon as conditions were met. In principle this seems to indicate that actual farmers behaviour (and decision to plant) has been influenced more than usual by non-rainfall related factors, such as access to agricultural Page 24

inputs, etc. Using the actual planting dates in Africa RiskView (instead of the planting criteria and windows defined during the customisation) results in a significant increase in the numbers of affected people. 4.4.3. Crop Cycle Length The reported crop cycle lengths appear to be slightly different from what has been currently customised, but switching to the reported values doesn t have a very significant impact on Africa RiskView s estimates. Assuming that the reported values are reliable, it would be safe to assume that these differences in LGP are not the reason why Africa RiskView didn t correctly pick up the magnitude of the drought (as was the case in Malawi). 4.4.4. Benchmark Switching from a 1-year benchmark to a 5-year benchmark is likely to result in a less volatile and more realistic benchmark. In terms of impact on modelled estimates, a reduction in the numbers of affected people can be noticed with a 5-year benchmark, which can be attributed to the overall good performance of the 2016 agricultural season. This reduction in the numbers of affected people is relatively consistent across the different regions. However, this doesn t seem to have contributed to Africa RiskView s moderate estimate of the drought severity. 4.4.5. Combination of explanations Assuming that there could be various explanations to the misrepresentation of the drought by Africa RiskView, the four approaches described above can also be combined in various ways 3 to explain the discrepancies observed between the model s estimates and the available ground data. In order to determine which combination constitutes a relevant explanation, both the individual approaches would have to be relevant, and the end result would have to match to some degree with the reference data. 3 However the rainfall correction cannot be combined with the benchmark change since it would require detailed station data for the past 5 years in order to apply the rainfall correction to each of the past 5 years before the final benchmark is calculated. Page 25

Annex I Kayes Koulikoro Segou Sikasso Mopti Tombouctou Gao BM = Benchmark; SW = Sowing Window; LGP = Length of Growing Period Cadre Harmonisé - Original (Pool 4) Original + BM Rain SW SW + BM LGP Nov 2017 (Ph. 3-5) Population % of Pop 2017 % of Pop 2017 % of Pop 2017 % of Pop 2017 % of Pop 2017 % of Pop 2017 % of Pop Bafoulabé 9,106 3% 6,380 2% 21,604 7% 67,130 23% 0 0% 0 0% 6,380 2% Diéma 27,512 10% 48,512 18% 25,632 10% 63,108 24% 49,284 18% 27,993 10% 59,174 22% Kayes 13,334 2% 0 0% 0 0% 63,530 10% 7,167 1% 0 0% 0 0% Keniéba 0 0% 0 0% 0 0% 0 0% 20,888 8% 21,374 9% 0 0% Kita 11,238 2% 48,433 9% 54,956 10% 93,323 17% 81,296 15% 82,689 15% 40,050 7% Nioro 20,819 7% 94,518 33% 11,340 4% 100,817 35% 94,826 33% 12,551 4% 94,518 33% Yélimané 4,586 2% 9,706 4% 5,174 2% 10,547 5% 10,016 4% 7,674 3% 9,706 4% Banamba 22,335 9% 0 0% 0 0% 0 0% 16,554 7% 18,318 8% 0 0% Dioïla 0 0% 0 0% 0 0% 0 0% 79,874 13% 81,188 13% 0 0% Kangaba 1,304 1% 0 0% 0 0% 0 0% 0 0% 23 0% 0 0% Kati 62,155 5% 0 0% 0 0% 0 0% 133,423 11% 134,290 11% 0 0% Kolokani 54,375 19% 8,972 3% 0 0% 0 0% 51,092 17% 49,051 17% 0 0% Koulikoro 2,736 1% 0 0% 0 0% 0 0% 26,345 10% 26,398 10% 3,410 1% Nara 9,429 3% 94,796 31% 89,764 29% 23,171 8% 102,648 34% 101,052 33% 87,498 29% Bamako 105,551 5% 0 0% 0 0% 0 0% 242,875 11% 246,489 11% 0 0% Barouéli 7,907 3% 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% Bla 0 0% 0 0% 0 0% 0 0% 0 0% 27,412 8% 0 0% Macina 6,134 2% 82,190 28% 15,299 5% 82,547 28% 20,247 7% 0 0% 86,676 29% Niono 9,481 2% 5,750 1% 11,707 3% 8,515 2% 0 0% 0 0% 543 0% San 4,334 1% 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% Ségou 18,088 2% 0 0% 0 0% 0 0% 0 0% 0 0% 123,295 14% Tominian 17,238 6% 0 0% 2,923 1% 0 0% 0 0% 0 0% 26,729 10% Bougouni 0 0% 0 0% 0 0% 0 0% 24,792 4% 24,823 4% 0 0% Kadiolo 6,325 2% 0 0% 0 0% 0 0% 10,882 4% 10,666 3% 0 0% Kolondiéba 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% Koutiala 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% Sikasso 9,549 1% 0 0% 0 0% 0 0% 19,900 2% 29,670 3% 0 0% Yanfolila 0 0% 0 0% 0 0% 0 0% 26,570 10% 26,593 10% 0 0% Yorosso 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% Bandiagara 36,646 9% 0 0% 0 0% 122,799 31% 116,508 29% 117,752 30% 0 0% Bankass 3,439 1% 0 0% 0 0% 106,988 32% 87,415 26% 70,384 21% 0 0% Dienné 37,902 14% 0 0% 0 0% 122,920 47% 0 0% 0 0% 0 0% Douentza 25,629 8% 0 0% 0 0% 143,646 46% 0 0% 0 0% 0 0% Koro 23,550 5% 0 0% 0 0% 154,969 34% 0 0% 0 0% 0 0% Mopti 57,504 12% 0 0% 0 0% 74,131 16% 103,725 22% 103,725 22% 0 0% Tenenkou 44,443 22% 0 0% 0 0% 108,386 53% 2,858 1% 7,941 4% 0 0% Youwarou 9,868 7% 0 0% 0 0% 66,706 49% 0 0% 0 0% 0 0% Diré 5,701 4% 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% Goundam 19,668 10% 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% Gourma-Rharous 10,101 7% 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% Niafunké 13,681 6% 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% Tombouctou 18,203 11% 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% Ansongo 13,723 8% 7,197 4% 0 0% 7,197 4% 7,197 4% 0 0% 7,197 4% Bourem 10,589 7% 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% Gao 28,026 9% 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% Ménaka 8,495 12% 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% Kayes 86,596 3% 207,549 8% 118,706 5% 398,456 16% 263,478 10% 152,282 6% 209,829 8% Koulikoro 152,335 5% 103,768 3% 89,764 3% 23,171 1% 409,935 13% 410,319 13% 90,908 3% Bamako 105,551 5% 0 0% 0 0% 0 0% 242,875 11% 246,489 11% 0 0% Segou 63,182 2% 87,940 3% 29,929 1% 91,062 3% 20,247 1% 27,412 1% 237,243 8% Sikasso 15,874 0% 0 0% 0 0% 0 0% 82,144 2% 91,753 3% 0 0% Mopti 238,981 9% 0 0% 0 0% 900,546 35% 310,506 12% 299,802 12% 0 0% Tombouctou 67,354 8% 0 0% 0 0% 0 0% 0 0% 0 0% 0 0% Gao 60,833 12% 7,197 1% 0 0% 7,197 1% 7,197 1% 0 0% 7,197 1% Total 790,704 4% 406,454 2% 238,399 1% 1,420,431 8% 1,336,381 7% 1,228,057 7% 545,177 3% Page 26