Reviewing our early-access pilot cohort in Sul de Minas across the 2025 harvest cycle, we found that the plot-level signal most consistently linked to yield shortfall was not the variable we had emphasized most before the season began. NDVI readings from May 2025, about six weeks before the region's peak harvest, matched actual per-hectare delivery results more closely than the two-season prior yield averages that then served as our main indicator of production stability.
This type of result comes from a cohort with actual growers and actual outcomes. Back-tested models show which inputs historically tracked yield in a training dataset. Live pilot results show whether those relationships remain valid in real time and at plot level. The 2025 harvest produced our fullest real-world calibration dataset since the early-access program started.
Here we explain what we found, how it changed our variable weighting, and what our cohort data cannot establish.
Cohort Profile and Measurement Approach
Our Sul de Minas early-access pilot cohort is concentrated in municipalities inside the designated Sul de Minas Altitude growing origin zone. Most plots sit between 800 and 1,200 meters above sea level. Cohort farms range from 4 hectares to 28 hectares, with a median of approximately 11 hectares. Every participant delivers to a registered cooperative, allowing us to use verified delivery records to measure outcomes.
For the 2025 harvest review, we recorded three outcomes for each plot: total bags delivered per hectare to the cooperative, cherry quality grade at weigh-in as reported by the cooperative, and whether the grower followed the projected repayment schedule. We assessed those outcomes against four inputs collected before harvest: the prior two-season average yield per hectare, plot altitude, soil classification, and the May 2025 NDVI reading.
These comparisons indicate direction rather than exact measurement. The cohort is large enough to identify patterns, but not to produce statistically dependable regression coefficients at high confidence intervals. We use the findings as qualitative calibration input, not as a published research dataset. They describe our cohort observations and do not represent claims about the wider Sul de Minas grower population.
What the NDVI Timing Showed
The clearest pattern in the 2025 results was the predictive value of May NDVI readings, taken approximately six weeks before peak harvest. In Sul de Minas's standard Arabica coffee growth cycle, May marks the shift from fruit development, which runs from January through April, into cherry maturation. A healthy May canopy indicates successful fruit set and early cherry fill, before major stress or disease pressure may arise during maturation.
Within our cohort, plots with above-median May NDVI readings achieved above-median per-hectare yields at considerably higher rates than plots with below-median readings. The relationship was not exact, and individual plots diverged in both directions. Still, the directional signal exceeded our expectations and was stronger than the two-season historical average would have suggested for those same plots.
This finding affects underwriting because it points to a more forward-looking input than historical averages by themselves. A grower seeking credit in April or May can have that year's May NDVI reading added to the production risk score before harvest, providing a current view of expected seasonal yield instead of depending fully on prior-season history.
Following the 2025 harvest review, we increased the relative weight of the most recent 60-day pre-harvest NDVI reading in the production risk composite. Prior-season yield history remains included, but its relative weight is now slightly lower than it was before the 2025 calibration.
Weather Index Patterns in INMET Data
We obtain weather index data from INMET (Instituto Nacional de Meteorologia) for the microclimate zones that cover our cohort plots. Two weather patterns stood out when we reviewed yield outcomes from the 2025 harvest.
First, flowering in September through October 2024 was followed by uneven rainfall in November through December 2024. Several two-week dry spells occurred during early fruit set. Plots with stronger soil water retention, as indicated by the soil classification input, maintained more stable NDVI readings during that period. This matched the expectation that latosol soils with more organic matter cushion short-term moisture stress better than thinner soil profiles.
Second, April 2025 weather records showed below-normal minimum temperatures in two altitude zones within the cohort, approaching zero Celsius at the highest plots. We marked this in our monitoring at the time. Cooler conditions during late fruit development seemed to slow maturation in those plots, but did not create the severe damage associated with actual frost. The affected plots delivered later in the harvest window without lower per-hectare yields. In some cases, the longer development period coincided with higher cup quality scores at cooperative weigh-in.
No frost event caused production damage in our cohort area during the 2025 season. We mention this because frost is a genuine and historically important risk in Sul de Minas, yet 2025 did not provide a statistically useful test of the model's frost-scenario handling.
How Soil Types Varied Across Yields
The soil variable followed the direction we expected, though its variation was broader than the model's prior weighting implied. Plots with latosol vermelho-amarelo achieved median or above-median yield rates more often than plots with less favorable soil profiles. However, variation within each soil category was considerable, reflecting management differences that soil classification by itself cannot show.
The most helpful adjustment concerns not the soil categories themselves, but their combination with NDVI. A latosol vermelho-amarelo plot with below-median May NDVI likely points to a management or seasonal problem rather than a soil constraint. In that application, the composite score should give the NDVI signal more influence than soil type. In contrast, a plot on less favorable soil with strong May NDVI may indicate unusually good management or a particularly favorable microclimate, so NDVI should likewise carry more influence.
In our current model development iteration, we are testing an interaction term for soil type and NDVI. The 2025 cohort does not contain enough observations to confirm that interaction with confidence, so it remains experimental.
Repayment Results and Model Performance
Through the 2025 harvest cycle, most cohort borrowers met their repayment schedules as arranged. Growers scoring above 65 in our production risk model performed as expected. The intermediate range, from 40 to 65, varied more. Some borrowers requested the structured hardship extension available to growers whose harvests fall below projections. None of those extensions became defaults. The harvest-cycle repayment structure allowed growers to change timing while continuing to service the debt.
We are not releasing specific cohort default rates or repayment percentages because the cohort is still early-stage, and the figures are not yet stable enough for public reporting without creating a misleading impression either way. We can report that directional performance matches our model targets. Cases needing restructuring were concentrated in the lower scoring range, which is the expected pattern when the model is functioning as intended.
What the Data Can and Cannot Show
We want to be clear about the boundaries of these findings. Our cohort remains concentrated geographically in Sul de Minas, focused on altitude, and still expanding. The patterns seen in 2025 may differ in Cerrado Mineiro or Chapada de Minas, where growing conditions, altitude profiles, and cooperative infrastructure are not the same as in Sul de Minas.
A single harvest cycle cannot clearly distinguish calibration signal from noise. Biennial bearing means an odd-year harvest cohort may produce different results from an even-year cohort. We need at least two complete biennial cycles of cohort data before making confident statements about which model weights remain stable and which reflect conditions specific to a year. The 2025 results changed our NDVI timing weights, and we expect to review them again after the 2026 harvest as more data becomes available.
The point of the 2025 harvest review is not to claim that the model was proven right. Its value is in giving us real-world results to compare with model predictions, showing that one input, NDVI timing, was stronger than expected, and confirming that the general structure of production-based credit scoring is following the actual risk patterns in our cohort. That is what early-stage calibration data should do: offer directional guidance rather than final answers.