In April 2016, the government of Malawi declared a national emergency. The rains had failed across southern Africa for the second year running, and 6.7 million Malawians were food-insecure, unable to feed themselves until the next harvest.
This was the scenario Malawi had insured against. For the 2015/16 season, the government had paid almost US$5 million for a sovereign drought policy from ARC Ltd, the insurance affiliate of the African Risk Capacity, the African Union's disaster risk pool. The policy was parametric: instead of assessing losses on the ground after a disaster, it watches an index. Satellite rainfall estimates feed a crop model, the model estimates drought response costs, and when that estimate crosses a threshold, money moves automatically. No loss adjusters, no months of claims paperwork. Speed is the whole point.
The drought came. The emergency was declared. The model concluded that no payout was warranted.
The crop that wasn't there
The explanation most widely reported in the press, chiefly in Bloomberg's 2024 feature on Malawi and in coverage of ActionAid's 2017 study, is a single calibration choice. Africa RiskView, ARC's modelling platform, had been customised to track long-cycle maize. Most Malawian farmers had switched to planting short-cycle varieties. In the model, the long-season crop rode out the erratic rains. In the fields, the crop actually planted did not.
The press accounts add a detail that makes this worse than a transcription error. Short-cycle hybrids are reported to be more sensitive to drought during flowering, not less. So the model was not merely watching a different crop. It was watching the hardier one while the fragile one stood in the field. Of the available ways to be miscalibrated, this one ran in the worst direction.
One attribution note, because it matters: the maize-cycle account rests on press reporting. The 75-page independent evaluation of ARC, commissioned after the crisis, never uses those terms. What the evaluation found instead is less tidy and considerably more useful.
What the evaluation actually found
The evaluation (by the e-Pact consortium, led by Oxford Policy Management, October 2017) starts where every parametric contract starts, with basis risk: "the risk that there will be a mismatch between the payout that is triggered and the actual situation on the ground." Basis risk is not a defect in parametric insurance. It is the price of the speed. Pay on an index and you accept that the index can diverge from the loss.
What turns an accepted risk into a scandal is everything around it, and on that the evaluation's Malawi case study is specific.
The calibration process was real, and it still missed. ARC describes customisation as a year-long process run through national technical working groups, drawing on local expert knowledge. Africa RiskView feeds rainfall estimates into the Water Requirements Satisfaction Index, "an operational crop model originally developed by the United Nations Food and Agriculture Organization," using "information about crops, such as soil and cropping calendars," and countries select their own risk-transfer parameters. This was not a default nobody read. It was a deliberate, expert-staffed fitting of model to ground, and the evaluation's verdict is that it "does seem to have neglected adequate input from agronomists, agro-meteorologists and other critical expert stakeholders, and appears to have been too removed from the 'ground'."
The check that could have caught the gap ran after the season. Ground-truthing exercises in three districts in April and May 2016 "revealed discrepancies between ARV and realities on the ground." Within a week, ARC commissioned a consultant from Malawi's Centre for Agricultural Research and Development to investigate why the model had failed. The check existed and it worked. It was simply scheduled for the moment its answer could no longer help anyone.
That post-mortem was never published. The findings "were disseminated via a workshop for stakeholders in Malawi, but the study itself was not made public." The one document that could confirm, refine, or complicate the maize story at source is unavailable to anyone who was not in the room.
Nobody had been required to understand the failure mode in advance. "ARC does not have a 'plain English' basis risk policy, and does not go through a formalised process to document and record agreement with countries that they understand the concept." A government spent $5 million on an instrument whose one catastrophic failure mode, a real drought with no payout, it had never been asked to demonstrate it understood.
And the causes are genuinely contested. The evaluation says its case study "suggests a more complex and nuanced story" than the single-parameter version. Opinions on what went wrong "vary widely, largely according to the Ministry and institution to which respondents belong," and some stakeholders pointed to a convergence of factors including prior years' climate disasters, corruption, and policy decisions. The evaluation calls for "shared responsibility of model failure between ARC and the government." An honest account has to carry all of that: the clean story is press-sourced, the accountable document declines to endorse it, and the underlying technical report is unpublished.
What is not contested is the outcome. After public uproar and months of dispute, ARC agreed in November 2016 to pay, and $8.1 million reached Malawi in January 2017, nine months after the declaration of emergency, from a product sold on speed. The government's drought response cost about $395 million. ActionAid's verdict that May: the insurance "failed to deliver on its promise of timely assistance, sorely needed by 6.7 million food-insecure Malawians, due to major defects in the model, data and process used to determine a pay-out." Malawi stepped back from ARC afterwards. So did Kenya, which had its own concerns about whether the model had been adequately customised for its risk pools.
The model did not fail to measure the drought. It measured a different field, accurately, and reported the truth about a place that no longer existed.
You have already built one of these
A parametric instrument pays on an index, not on your loss. Read that definition again and it stops being about insurance. An eval suite pays out, by shipping the model, on a benchmark score, not on behaviour in users' hands. An SLO pays out, by silencing the pager, on measured latency, not on user experience. A KPI dashboard pays out, by ending the argument, on the metric, not on the business. Anyone who has shipped a metric that gates a real decision has written a parametric policy, usually without either party noticing the contract.
That is why this case deserves a working engineer's attention. The failure has a specific shape, and no part of it is "the model was inaccurate":
The instrument was calibrated correctly and became wrong. The farmers switched varieties. Nothing in the model degraded; the referent moved. Your eval set was representative the quarter you built it.
The mismatch was invisible from inside. Africa RiskView read real rainfall and ran a real crop model. There is no internal signal for "the thing I am scoring is not the thing that exists." That fact lives outside the instrument, and only ground contact reaches it.
The ground-truthing was scheduled for after the season. The check that found the gap ran when its answer was useless. Most eval refreshes are timed to the release calendar, not to the rate at which the world drifts.
Nobody had to sign that they understood the failure mode. No plain-English statement of what the index cannot see, agreed before the premium changed hands. Ship a dashboard and ask who has ever been required to state, in writing, what it is blind to.
The post-mortem stayed in the room. The explanation of the failure went to a workshop and stopped. Everyone negotiating the next contract in the same instrument class does so without it.
This is also a different disease from the insurance failure that usually gets the attention, correlated risk, where one shared dependency fails many insureds at once and threatens the pool itself. Correlation breaks the insurer. Basis breaks the promise: the loss was real, the index looked away, and the contract performed as written.
The season after
Malawi came back. In July 2025, ARC announced a payout of $3,376,783 to Malawi for the 2024 drought: about $3.1 million on the traditional sovereign policy and $311,037 on a newer anticipatory policy, which ARC describes as allowing "early response to a failed planting season." Paying on the planting season rather than after the harvest is the index moving closer to the field, and that is what learning would look like.
Whether the 2016 lesson was absorbed is harder to establish. The 2025 announcement makes no reference to that season and describes no change to how the model is customised or ground-truthed. That is a claim about one press release, not about ARC's practice; nine years is a long time, and much may have changed. But the document that would show it, the technical post-mortem of the season everyone still argues about, remains unpublished.
Sources
Independent Evaluation of the African Risk Capacity (ARC), Formative Phase 1 Report, e-Pact consortium (Oxford Policy Management with Itad), 27 October 2017.
African Risk Capacity, "Africa RiskView," the operator's published description of its customisation process and crop model (arc.int), accessed August 2026.
African Risk Capacity, "Government of Malawi Receives Insurance Payout from ARC Group Following 2024 Drought," press release, 25 July 2025.
ActionAid, The Wrong Model for Resilience, 24 May 2017.
Bloomberg, 2024 reported feature on Malawi and ARC's parametric drought cover (paywalled).
If you ship a metric that gates a real decision, the fifth failure mode above is the one you can actually fix this week: the explanation of why an instrument was trusted, and what it could not see, usually stays in the room. Chain of Consciousness records that reasoning as a durable artifact, so the next person negotiating against the same index can read why it was believed rather than reconstruct it.
pip install chain-of-consciousness
npm install chain-of-consciousness
Top comments (0)