The Honest Forecast Has a Date on It

A forecast becomes useful when it says what could prove it wrong, when the answer will be known, and who will publish that answer.

The forecasts discussed here were recorded on 16 October 2026. The observations below use information available by that date; outcomes remain unresolved.

A signal is not an outcome

On 16 October, Kenya’s September inflation rate was known: 6.8 per cent year on year, with food inflation at 9.5 per cent and transport inflation at 15.6 per cent. Those are measurements of prices already observed, not a guarantee about December. The distinction is elementary and easy to forget when the recent number is frightening. Kenya National Bureau of Statistics, September 2026 CPI release

The international food signal was different. The Food and Agriculture Organization’s September Food Price Index averaged 136.0 points, 1.5 per cent above August and 5.8 per cent above the same month a year earlier. It tracks a basket of globally traded commodities. It is not Kenya’s retail food index, and it cannot tell us how much a household in Kisumu or Nairobi will pay for maize flour. FAO, “Food Price Index rises in September,” 2 October 2026

The climate signal had another source of uncertainty. NOAA’s 8 October discussion put the chance of an exceptionally strong El Niño—defined there by a three-month relative Oceanic Niño Index of at least 2.5°C—at 83 per cent for October–December. That is a forecast for a Pacific climate index, not an 83 per cent chance that any named Kenyan river will flood. Regional rainfall depends on more than one ocean index, and local impacts also depend on soils, drainage, settlement patterns and preparedness. The World Meteorological Organization likewise says El Niño’s rainfall effects differ by region. NOAA Climate Prediction Center, ENSO Diagnostic Discussion, 8 October 2026; WMO, Global Seasonal Climate Update for August–October 2026

These are observations and forecasts, not interchangeable evidence. A rise in the global food index can raise concern without specifying the Kenyan price response. A stronger El Niño shifts the range of weather planners should consider without naming a flood, date or casualty count. Forecasting starts by respecting those distinctions.

Put the event in a sentence

A forecast should be a question that an independent reader can answer yes or no. It needs a threshold, a deadline, and a source that determines the outcome. “Food prices will get worse” is not a forecast. “FAO’s December 2026 Food Price Index will be above 136.0 points when published” is. So is “Kenya’s December year-on-year CPI inflation will be at least 7.0 per cent in the KNBS release.” A forecast without a resolution rule can be reinterpreted after the fact to protect its author.

The October 16 forecast register included three questions tied to the signals above: whether Kenya’s December CPI would reach 7 per cent year on year; whether FAO’s December index would exceed its September level of 136.0; and whether Kenyan authorities would report at least 50 flood-related deaths during the October–December short rains. The probabilities were 55, 60 and 45 per cent respectively. They were subjective, uncalibrated judgments recorded before the outcomes were known, rather than probabilities derived from September’s inflation or NOAA’s climate forecast.

That separation matters because a prediction can be poorly framed even if its author sounds confident. “A major flood will happen” leaves “major” undefined. Does the question mean rainfall above a threshold, a national emergency declaration, deaths, homes destroyed or crops lost? Each captures a different consequence. The outcome source must also be named: a meteorological dataset, an official disaster report, or a defined newswire standard. When evidence is mixed or delayed, a forecast should say how it will be resolved rather than choosing the most favorable interpretation afterward.

Score the probability, not the confidence of the prose

For a yes-or-no event, the Brier score is the squared difference between the forecast probability and the outcome: (p − y)², where yes is 1 and no is 0. A 60 per cent forecast that happens scores 0.16: (0.60 − 1)². If it does not happen, it scores 0.36: (0.60 − 0)². Lower is better. A forecaster who assigns 50 per cent to every question scores 0.25 each time, whatever the answer.

One score does not reveal skill. A forecast at 60 per cent will miss four times in ten in the long run; a single miss does not make it foolish. Calibration asks whether events assigned roughly the same probability occur at roughly that frequency across many resolved questions. Discrimination asks whether higher probabilities tend to be assigned to events that happen more often than those that do not. A forecaster can be well calibrated by issuing cautious probabilities for everything and still fail to distinguish a real warning from noise. Both qualities require a track record across enough questions. Gneiting and Raftery, “Strictly Proper Scoring Rules, Prediction, and Estimation,” Journal of the American Statistical Association, 2007

The scoring rule also discourages false certainty. Assigning 99 per cent to a disaster and missing it is costly. Assigning 1 per cent and seeing it happen incurs the same squared error: 0.9801. Good forecasts do not avoid commitment; they make their degree of commitment visible and revise it when relevant evidence arrives. A forecast’s probability should change when the evidence changes, not merely when its author is embarrassed.

Base rates are a disciplined starting point. Before focusing on a dramatic new signal, a forecaster can ask how often a comparable event has occurred in a relevant class of cases. That history cannot decide the next outcome: rainfall, infrastructure, prices and policy all change. It does force the analyst to explain why this case should differ from the ordinary pattern instead of treating every headline as unprecedented.

A probability is not an instruction

The same forecast can justify different actions for different decision-makers. A county that can clear a drain at low cost may do so at a modest flood probability. Relocating residents, closing a bridge or moving a clinic’s supplies has larger costs and needs more local information. The decision depends on the probability, the severity of the loss, who bears that loss and the cost of preparation. No percentage by itself answers those questions.

One simple decision rule compares expected loss. If an action costs KSh 100,000 and would prevent KSh 1 million in damage if a defined flood occurs, the break-even probability is 10 per cent, before accounting for distributional effects and uncertainty in either cost. That is not a flood forecast. It is a transparent way to discuss whether an action may be worthwhile. In practice, avoidable deaths and unequal ability to recover make the decision more complicated than a single cost ratio.

The counterargument is real: teams can drown in alerts. A warning with no owner, deadline or decision attached becomes noise. The remedy is not to publish more decimal places. It is to choose a short list of indicators, assign who checks them, decide what evidence changes the action, and review whether the action helped. A probability is most useful when it changes a decision without pretending to settle the future.

For a clinic exposed to short rains, the operational questions are concrete: Does the site flood? Can medicines be raised above likely water levels? Is the referral route still usable if a road closes? Who informs the county health team, and what trigger moves supplies or patients? For a food trader, the relevant question may be whether transport routes and local stocks can absorb an import delay. A national index identifies a pressure; local logistics determine how it reaches a particular service or household.

A forecast needs humility and a receipt

Good practice is modest but demanding: state the event, probability, deadline, source and resolution rule before the answer is known; keep the original forecast visible; record updates with reasons; then score it whether the outcome flatters the forecaster or not. If the question was ambiguous, revise the question design next time rather than quietly changing what the old forecast meant.

An update should preserve the first estimate and say what changed. A revised probability is useful only if readers can distinguish new information from a new interpretation of old information. The trail matters: the publication date, the indicator that moved, the source behind it and the reason the odds changed. Without that record, a forecaster can sound right after the fact while learning nothing from the miss.

On 16 October, the September inflation print, September FAO index and NOAA’s El Niño discussion were all useful signals. None dictated the next quarter. The flood probability was not a weather report; the inflation probability was not a promise about the household basket; the global food index was not a Kenyan price forecast. A forecast earns trust by showing where evidence ends and judgment begins.

The honest forecast has a date on it because a prediction without a deadline can survive every outcome. The date gives the public a way to check it, the resolution rule makes the answer harder to move, and the score records what the forecaster learned. Uncertainty does not excuse vagueness. It is the reason to be precise.

Evidence cutoff for the signal snapshot: 16 October 2026. The three probabilities above are unresolved forecasts as recorded on that date.

The rain gauge on a Davis Instruments wireless Vantage Pro2 Plus weather station at Ladd Observatory, photographed 4 February 2008 in Providence, Rhode Island. Historical scientific equipment illustrating measured observations; not a Kenyan weather station or a current forecast. Credit: Michael L. Umbricht / Ladd Observatory, via Wikimedia Commons. Image source · CC BY-SA 4.0 International