Reference

Weather impacts by growth stage

Knowing that Argentine corn silks in January does not tell you a hot January is bearish production. That is a separate claim, it has a mechanism, and in a small number of cases we have actually tested it. This page keeps the three apart: timing, mechanism, evidence.

For the timing half, see the crop phenology calendar.

The evidence tier — read this before quoting anything below

Every impact row carries a tier. The tier is the whole point of the page; an untiered weather claim is the failure mode this entry exists to prevent.

tiermeaningwhat you may say
T1 — validatedFalsified against a named production series in mcp_server/series_loaders.py PHENOLOGY: correlation, p-value, and neighbouring offsets shown to collapse.May be quoted with its statistic, and only alongside the outcome series it was tested against.
T2 — modeledWe compute an index for it with published, cited thresholds (frost, leaf rust, dry-hot-wind, harvest rain, THI). The mechanism is from literature; the classification is ours and runs on our data.May be quoted as a risk classification, never as a yield number.
T3 — documentedEstablished agronomy with a named reference event we can point to in our own record. Not tested here.Quote as mechanism plus the reference year. Never attach a magnitude.
T4 — indicativePlausible, directionally standard, region-general. The weakest thing on this page.Use as context only. Say "typically".

Only nine rows on this page are T1. They are marked, and they are the only place a number like "r = −0.82" belongs. Everything else is mechanism, not measurement. Blurring that reopens exactly the free-form-window multiplicity problem the registry exists to prevent — see the warning in verify_phenology_calendar.py.

The general law, stated once

Three rules explain most of the table and are worth holding in preference to memorising it:

1. Sensitivity is not flat across the season. The same anomaly is noise in one stage and the story in the next. The ordering is nearly universal: flowering ≈ fill > vegetative > sowing > maturation > dormancy. 2. Reproductive stages fail on YIELD; harvest fails on QUALITY. These are different trades. A yield failure removes tonnes and moves the flat price. A quality failure moves

differentials, grades and certified stocks while national tonnage barely changes — Vietnamese harvest rain and US SRW vomitoxin are quality events, not supply events, and calling them supply events is a recurring error. 3. The damage function is one-sided and it flips. Water is a deficit risk in every stage except sowing and harvest, where it becomes an excess risk. Any statement of the form "rain is good for X" is wrong for at least two months of X's own calendar.

Impact matrix — stage × anomaly

The shared stage vocabulary is the one fixed in crop phenology calendar. series names the column that would show it in our data.

sowing — establishment

anomalyeffectmechanismtierseries
Excess rain / saturated soilThe dominant risk in this stage. Delays or prevents planting; forces acreage switching or prevent-plantMachinery cannot travel; seed rots in anoxic soilT3 — France 2016 and 2024, the two worst French wheat crops in decades, were both excess-rain years and both rank as the wettest May–Jun in our recordmms, soil_moisture_l1
Deficit / late rain onsetDelays sowing; in rain-gated systems this compresses the whole downstream calendarSowing is gated on the wetting front, not the dateT3 — Matopiba soy sowing waits on the rain onset; Australian sowing waits on the autumn breakmms, cwsi_z
Cold soilSlow, uneven emergence; replant riskGermination is temperature-rate limitedT4avg_temp_c, min_temp_c
HeatLargely irrelevant at this stageT4

Read the sign backwards here. A wet anomaly at sowing is a risk reading, and our CWSI convention (positive = drier, see cwsi methodology) will show it as a benign negative. Sowing is one of the two stages where a "good" CWSI is bad news.

vegetative — canopy build, tillering

anomalyeffectmechanismtierseries
DroughtReduced tillering / leaf area — caps the yield ceiling but is often recoverableFewer tillers and less intercepted radiationT3cwsi_z, soil_moisture_l2
HeatAccelerates development, shortening the seasonFaster thermal-time accumulation, less total interceptionT3avg_tmax_c, TSA (thermal stress accumulation)
Excess rainNitrogen leaching; disease establishmentT4mms

This stage is where markets over-react. Vegetative stress is the most visible (bad crop-condition ratings, brown fields, drought-monitor expansion) and the most recoverable. A June corn-condition slide that resolves before silking has historically cost far less yield than the ratings implied.

flowering — anthesis, pollination, blossom

The shortest and least forgiving window on the page. Damage here is not recoverable: the organ count is set and cannot be rebuilt.

anomalyeffectmechanismtierseries
HeatPollen sterility, kernel/floret abortionPollen viability collapses above roughly 35 °C for maize and 32–34 °C for wheat; the failure is at the daily maximum, not the meanT1 for Russia/Ukraine and Australia (see below); T3 for US cornmax_tmax_c, avg_tmax_c
DroughtSilk–pollen desynchrony in maize; flower abortion in soySilk emergence delays while pollen shed does not, so shed finishes before silks are receptiveT1 Argentina corn silking soil moisture r = +0.74 (differences) vs Argentine corn productioncwsi_z, soil_moisture_l2
FrostTotal loss of the affected floretsIce nucleation in reproductive tissueT2 — modeled for coffee (coffee frost risk) and Chinese NE (china crop weather methodology); T3 for Australian east wheat, where frost at flowering is the documented riskmin_temp_c, frost model
Excess rainPoor pollination; blossom drop; in coffee, blossom washingPollen is washed off / dilutedT3mms
Dry spell then rain (coffee only)This is the trigger, not the damage. The bloom needs a dry period followed by a rain event to set synchronouslyWater-deficit release triggers anthesisT1 Brazil arabica flowering r = −0.59, p = 0.021 vs CONAB arabica production, at offset −1mms, cwsi_z

Coffee flowering is the one stage on this page that sets the NEXT crop year, not the current one. The offset is −1 and it fails silently if you get it wrong. A September 2026 bloom failure is a 2027 supply story.

fill — grain fill, pod fill, coffee granação

The yield-determining window for every crop here. Damage is partly recoverable in extent (fewer, lighter grains rather than none) but this is where most of the variance lives.

anomalyeffectmechanismtierseries
HeatShortened fill duration → shrivelled, light grain; test weight fallsAbove ~30–32 °C the fill rate rises less than the fill period shortens, so total assimilate fallsT1 — the strongest results we have. Russia grain-fill tmax r = −0.82, p = 0.0002; Australia east tmax r = −0.80, p = 0.0006; Australia west r = −0.79; Argentina soy pod-fill tmax r = −0.60avg_tmax_c, max_tmax_c, TSA
DroughtAborted grains, reduced sizeAssimilate supply and remobilisation both limitedT1 — Argentina soy pod-fill precip r = +0.65, p = 0.006; Ukraine precip r = +0.56cwsi_z, mms, soil_moisture_l3
Terminal heat (wheat, subtropics)The named failure mode for Indian rabi wheatA late-season heat spike truncates fill on an already short seasonT3 — March 2022, the hottest March in our record (+2.58 °C), followed by an export ban within weeks. ⚠ The India window is registered but unconfirmed (~93% irrigated, production is near-pure trend, lag-1 autocorrelation 0.94)avg_tmax_c
Veranico (Brazil)A mid-summer dry spell inside an otherwise wet seasonTwo to three rainless weeks during pod fill on soils with little bufferT3 — the Matopiba window is registered but unconfirmed against national production, because Brazil's national soy variance is driven by SOUTHERN droughtscwsi_z, mms
Excess rainMostly benign; disease pressure risesT4mms

Heat and drought are not separable in this stage and should not be reported as two findings. They co-occur (the same blocking pattern produces both), and our CWSI already mixes them by construction — it is (precip − ET) / soil, and ET rises with heat.

maturation / dry-down

anomalyeffectmechanismtierseries
RainDelays dry-down; raises harvest moisture and drying costT4mms
Early frostHalts fill before physiological maturity — light, immature grainKills the canopy while the grain is still fillingT2 — modeled for the Chinese Northeast (Aug–Sep window, china_frost_model.py); T3 for the US Northern Plainsmin_temp_c, frost model
HeatAccelerates dry-down; mildly positive for harvest logisticsT4avg_tmax_c

harvest — a QUALITY window, not a yield window

anomalyeffectmechanismtierseries
RainThe dominant risk. Grade loss, sprouting, mycotoxin, colour and cup defectsWet grain sprouts in the head; wet coffee cherry ferments on the patioT2 — modeled for the Chinese NCP (china_harvest_rain_model.py, per-province May 15 – Jun 25 windows, sprouting); T3 elsewheremms, rain_days
Excess rain, coffeeCup quality and defect count, not tonnageDrying is interrupted; over-fermentationT3 — the Vietnamese Oct–Dec riskmms
DroughtBenign to helpfulT4
Wind / stormLodging; in Central America, hurricane exposure Jun–NovT4max_wind_ms

Do not translate a harvest-rain event into a production cut. It moves grade distribution, deliverable supply and certified stocks. The tonnage usually shows up; it just shows up as feed wheat instead of milling wheat.

dormancy — winter cereals only

anomalyeffectmechanismtierseries
Cold WITH thin snow coverWinterkill. Stand loss, requiring reseeding to a spring cropSnow is the insulator; the crown injury threshold is a crown temperature, which cold alone does not determineT3 — and this is a joint condition, never a minimum temperature alone. Our own normals make the point: January snow water equivalent averages 33.0 mm in the US Northern Plains against 3.3 mm in the Central & Southern Plains, so the HRW belt reaches its cold with almost no insulationmin_temp_c, avg_snow_swe_mm
Warm spell then hard freezeDe-hardening followed by injury; and at green-up, a spring freeze on jointed wheatCold-hardiness is lost within days of a warm spell and is slow to rebuildT3avg_temp_c, min_temp_c
DroughtLow sensitivity while dormant; matters at green-upT4cwsi_z

Snow water equivalent is not snow depth. avg_snow_swe_mm is the water content; the insulating depth depends on density and is not carried. A dedicated winterkill treatment is scheduled for late September 2026 and this row is deliberately thin until then.

Crop-specific notes that do not fit the matrix

Corn. The sensitivity peak is extraordinarily sharp — roughly the fortnight bracketing silking. US 50% silking has a median of Jul 18 with an observed range of Jul 9 – Jul 26 (us crop phenology observed), so "July heat" is not one event: a heat ridge in the first week of July hits a crop that is mostly still vegetative, and the identical ridge in the third week hits pollination. Date the ridge against the stage, never against the month.

Soybeans. The critical window is later than corn's and better defined by pod fill (median 50% setting pods Jul 31) than by bloom, because soy compensates: it flowers indeterminately over weeks and can reset pods after a stress. That compensation is why an August drought is worse for soy than a July one, and the reverse is true for corn.

Wheat. Three separate exposures that are routinely collapsed into one: autumn sowing conditions (can the crop be established?), winter survival (winterkill), and May–June fill (heat and drought). They are separated by six months and can point in opposite directions in the same market year.

Coffee. The only crop here with three critical windows in one calendar and a biennial bearing cycle on top. Arabica alternates high and low bearing years (lag-1 autocorrelation of production differences −0.84), which means any weather effect must be read against the expected phase of the cycle, not against last year's number. Conilon does not bear biennially (lag-1 +0.84 in levels, trend-dominated), which is precisely why the robusta windows inherited from arabica failed falsification.

Cotton. Still the worst-covered market we publish a report on, but the dating gap closed on 2026-09-06. PCT SETTING BOLLS and PCT BOLLS OPENING are now ingested back to 2000, at US-national level and for 15 states, so the boll-fill period can be dated: US-national 50% setting bolls Jul 28, 50% bolls opening Sep 14, and the window is the span between them — NASS publishes no cotton bloom series, so neither date alone names it. Squaring (Jul 05) is the flower-BUD stage, before anthesis, and marks the approach rather than the window.

What remains missing is the statistical half: cotton still has no registered window — nothing here has been falsified against a cotton production series, so a measured date gates when to look, not what it means. The mechanism rows above still apply — heat and moisture stress at flowering and boll fill shed squares and bolls — and the stack can now tell you the crop is in that stage on a given date.

Excess versus deficit — the asymmetry table

The single most common error is treating "wet" and "dry" as one axis with a good end.

stagedeficit riskexcess riskwhich dominates
sowingmoderate (gating)highexcess
vegetativemoderatelowdeficit
floweringhighmoderatedeficit — except coffee, where the timing of both is the signal
fillhighlowdeficit
maturationlowmoderateexcess
harvestnonehigh (quality)excess

How to use this from a machine client

The decision procedure, in order. Stopping early is the point — most anomalies should exit at step 2.

1. Locate the stage. US → us crop phenology observed (measured, with range). Everywhere else → crop phenology calendar (descriptive, treat boundaries as ±1 dekad). 2. Is the stage critical? If not, report the anomaly as context and stop. Do not escalate a −2σ dekad in dormancy. 3. Look up the anomaly in the matrix above to get direction and mechanism. 4. Check the tier. T1 → the statistic and its outcome series may be quoted. T2 → quote the risk class, not a yield. T3/T4 → mechanism only, no magnitude. 5. Check for a modeled index that already covers it — frost, rust, dry-hot-wind, harvest rain, THI. If one exists, it supersedes the matrix row, because it runs on thresholds rather than on adjectives. 6. State the observation date and the stage together. "Third dekad of July, corn at silking, tmax +2.4 σ" is a usable sentence. "Hot in July" is not.

Caveats & limitations

Change history