Blog When the rains don't follow the calendar

The data comes from CHIRPS data, an open-access product from the USGS/UCSB Climate Hazards Center that blends satellite imagery with ground-station readings to estimate rainfall across Africa. We used monthly totals for Kisumu, January 2018 to December 2022 resulting to 60 months. There's no steady rise or fall over the five years, but in most of the studied years, rainfall totals to over 350mm some months and under 50mm others.

In Kisumu, a very wet month can be followed just weeks later by a near-dry one, and while it's happening, there's no way to know if that's normal or the start of real trouble. That's why planting decisions, advisories, and planning can't rely on 'normal' rainfall alone; watching the season as it unfolds is what catches a problem in time to act.

Knowing how much rain a season brings isn't the same as knowing when it will arrive, and for farmers, timing is usually the harder problem. Research across the Lake Victoria Basin has documented growing intra-seasonal rainfall variability, delayed onset, and more frequent dry spells (Kenduiywo et al., 2026). The two views displayed in Figure 2 and Figure 3 below answer two different timing questions, from two different angles.

From Figure 2, each of the five small charts traces one year's rainfall month by month. This visual is useful for spotting a year, like 2021, that broke from the usual pattern. On the other hand, each panel in Figure 3 traces one calendar month across all five years, with an orange line marking that month's five-year average. This is useful for checking how much a given month's rainfall varies over the five years, e.g., does May look like May every year, or is every May a surprise?

Kisumu's long rains run March–May, and in four of five years, May sits at or near a seasonal peak (Figure 2). In 2021, rainfall in May collapsed to under 50mm, causing a sharp dip in the middle of the main rainy season. This is exactly the kind of anomaly a farmer relying on “normal” timing would have felt directly.

Seasonal and Trend decomposition using Loess (STL) decomposition splits the raw data into three parts including trend (the underlying long-term direction of the time series data; Osterrieder, 2023), seasonality (repeating patterns in the data, such as monthly or yearly patterns; Osterrieder, 2023), and remainder/noise (random fluctuations in the time series data that cannot be explained by the trend or seasonality; Osterrieder, 2023). Seasonal decomposition plots show more in-depth analysis of the data and identify patterns and trends that may not be apparent in basic plots.

Figure 4 shows all the three time series components while Figure 5 zooms into the seasonal piece alone, with all five years overlaid, to make its shape easier to read.

In Figure 4, the trend moves gently within a narrow band (roughly 140–255mm) with no long-term increase or decrease in rainfall across the five years. The remainder, which swings by more than 200mm with no obvious pattern, represents the part no seasonal calendar will forecast and the surprises layered on top of the normal rainy season. This is exactly the part a farmer still has to prepare for. One point in the remainder row is marked directly: May 2021's remainder is the single largest deviation anywhere in the five-year record, illustrating how an abnormal pattern looks like once trend and season are stripped away.

• September is the most dependable month on record; it is consistently dry with the least year-to-year variation.

• May brings the second-most rain on average, but it's also one of the most variable, ranging from under 50mm to over 350mm across the five years. This is exactly the kind of month farmers rely on most yet one of the hardest to forecast. This matches farmer-reported behavior in a 2025 baseline survey across Kisumu County. The survey found that irrigation activity for key crops is lowest in May (Wanjau et al., 2025), a farmer-reported indicator of the same seasonal risk pattern found in the rainfall record.

• Seasonal averages describe a typical year, but they don't describe a particular year. May 2021 is the clearest case - on paper, an average May looks like a safe planting month, but that year's actual rainfall broke sharply from it. If advisories lean only on long-term averages, anomalies like this go undetected until it's too late to act. Monitoring rainfall as the season progresses, not just checking it against the seasonal norm, is what gives you the lead time to flag a bad season early and adjust guidance to farmers before the damage is done.

Findings from this blog line up closely with T-CSIA's (Targeted Climate-Smart Irrigation Advisory Using Integrated Climate Indicators) own work. T-CSIA's historical index, built on 44 years of data, already identifies eastern Kisumu, especially the lower Nyando basin, as a persistent hotspot of irrigation-related water stress. This blog's independent analysis, using a different method and a more recent five-year window, arrives at a consistent picture, i.e., rainfall in this area follows a real seasonal pattern, but its timing within that season is genuinely hard to predict. May 2021 offers some concrete, dated, on-the-ground example of exactly the kind of risk T-CSIA's index is designed to flag.