The 30 to 45 day window: forecasting demand before it lands

Two horizons govern every weather-sensitive business, and they almost never line up. One is how far ahead the atmosphere can actually be predicted. The other is how far ahead you have to act to be in market when demand arrives. The first is short and capped by physics. The second is long, set by ad approvals, learning phases, and inventory lead time. Confuse them and you either commit on noise or arrive after the spike has passed.

Start with the physics, because it sets the ceiling. The atmosphere is a chaotic system: tiny errors in the starting state grow until the forecast is no better than climatology. Edward Lorenz formalised this in the 1960s: his 1963 paper described the "butterfly effect", and his later work put a hard limit on skilful deterministic weather prediction of roughly two weeks (Lorenz, 1963 and 1969, Journal of the Atmospheric Sciences). Modern work has put a sharper number on it. Zhang et al. (2019), in the Journal of the Atmospheric Sciences, estimate the intrinsic predictability limit of midlatitude weather at about 10 days, with current models capturing most of that potential and only a few days of further skill theoretically available even with a perfect model.

That is the science behind the practical rule operators already sense. Day-to-day temperature forecasts hold real skill out to about 7 to 10 days; precipitation, which depends on finer-scale structure, decays faster, holding to roughly 3 to 5 days (Bureau of Meteorology, forecast verification; Penn State METEO). The European model, generally the global leader, keeps its anomaly correlation above the 0.8 "skilful" threshold to around day 10, and only crossed reliably past 10 days with its late-2024 model upgrade (ECMWF, Quality of our forecasts, 2024). Australia's national model, the Bureau of Meteorology ACCESS system, sits on the same physics and the same horizon. The headline: a credible day-specific weather forecast is a one-week instrument, not a one-month one.

Why demand still moves a month ahead

If the forecast is only good for a week, how can anyone plan a month out? Because the thing you are forecasting is not the weather. It is demand, and demand has a slower, more predictable structure than the daily temperature does. Three layers sit on top of each other.

The first is seasonality, the climatological baseline. You do not need a forecast to know that cooling demand rises into the Australian summer (December to February) and heating demand into winter (June to August); that pattern repeats every year and is predictable months ahead. The geography matters: cooling degree days increase towards the north of the continent and heating degree days towards the south, so cooling demand spans most of Australia and dominates the north (Brisbane, Perth, the tropics), while heating demand concentrates in the temperate and alpine south (Melbourne, Canberra, Hobart) (Bureau of Meteorology, heating and cooling degree-day climatology). The honest caveat is that this is climate, not weather, and the two must be separated. A study on retail sales by Badorf and Hoberg (2020), summarised in the climate-control literature, and the methods note in "Accounting for Climate When Determining the Impact of Weather on Retail Sales" (ResearchGate, 2023) both make the same point: if you do not net out the normal seasonal baseline, you will falsely flag categories as weather-sensitive and over-claim the weather effect. The seasonal layer is reliable but it is not the edge; everyone has it.

The second layer is the lag structurebetween a weather signal and the demand it produces, and this is where forecasting earns its keep. The lags are not uniform. L.E.K. Consulting's Weather Index for Service Essentials (WISE) framework, a US method we apply to Australian regions (NSW, VIC, QLD, WA, SA), documents that HVAC shipments respond with roughly a one-quarter lag to degree-day accumulation, while roofing demand spikes in the same quarter as a storm and then falls for two quarters afterward (L.E.K. Consulting, WISE framework, US data). In Australia that storm trigger has its own calendar: severe hail and storm damage in eastern Australia cluster in spring and early summer (September to December, peaking November to December in NSW and south-east QLD), while the northern tropical storm and cyclone season runs November to April (Bureau of Meteorology, severe storms archive; hail climatology of greater Sydney and NSW). At higher frequency the response is even tighter: electricity-load research finds outdoor temperature one to twenty-four hours prior is the dominant predictor of cooling and heating load, with a cooling-degree-hour index explaining about 90% of variance (Yes Energy, load-forecasting research). So a heat event produces near-immediate emergency demand and a delayed replacement wave, and a credible model has to carry both.

The third layer is the part operators feel in their P&L: the time it takes to act. A new campaign has to clear approval, then pass through a learning phase before its cost per result stabilises; inventory and crews have to be staged. None of that is instant. So the operative horizon is not when demand peaks; it is when demand peaks minus your own lead time to be ready for it.

The forecasting question is not "what will the weather be". It is "what can I commit to today and still be right when the weather arrives".

The 30 to 45 day rule, and the math behind it

The widely cited practical rule for weather-driven verticals is to launch campaigns 30 to 45 days before the seasonal peak. In the Australian calendar that means cooling-season creative live by October ahead of the December to February summer peak, and heating-season creative by April ahead of the June to August winter peak (WebFX, home-services seasonal search analysis, US data; timing inverted for the Southern Hemisphere). That is not arbitrary. It is the sum of the lead times you cannot compress.

Put real numbers against it. The figures below are illustrative of typical paid-search timelines, but the arithmetic is standard.

Round up for slippage and you land at the 30 to 45 day window. Notice what carries the rule: the learning phase and the staging time, not the weather forecast. The forecast only sets how wide your safety buffer must be. Google's documented conversion-based learning phase is roughly one to two weeks, which alone accounts for a third of the window.

How the lag itself is measured

The defensible method for finding which weather signal leads demand, and by how long, is textbook and worth stating plainly because it sets honest expectations. You compute the lagged correlation between a weather feature and demand, then keep the lag that maximises it, controlling for the seasonal baseline so you are measuring weather rather than calendar.

Degree days are the workhorse feature. A cooling degree day accrues for each degree the daily mean temperature sits above an 18°C base; heating degree days accrue below it. The Bureau of Meteorology publishes Australian heating and cooling degree-day climatology on an 18°C base (it also maps 12°C heating and 24°C cooling thresholds), so 18°C is the metric base to use here, not the 65°F US convention (Bureau of Meteorology, degree-day climatology, 18°C base). They convert a messy temperature series into a single demand-relevant number, which is what makes the lag estimable in the first place.

The payoff for doing this well is measurable, and it is honest to say it is modest in some categories. Adding weather to demand forecasts has been shown to improve accuracy by about 20.2% for grocers, 12.2% for casual dining, and 3.3% for home improvement (ToolsGroup, weather and demand forecasting). The home-improvement number is the cautionary one: weather matters most where the purchase is impulsive and weather-coincident, and less where it is planned. A good model tells you which of those you are.

The operating calendar

Translating all of this into a calendar gives a business something it can run on. The blue months are when you commit and launch; the amber months are when demand actually peaks. The gap between them is the 30 to 45 day window made concrete.

Horizon sources: ECMWF (2024) and the Bureau of Meteorology / Penn State METEO for the daily-weather windows; the seasonal window is the climatological pattern, predictable from BOM degree-day history rather than from any single forecast.

This is also where the metric we care about, Profit Velocity, comes in. Getting the timing right is not about spending more; it is about converting the same budget into durable contribution margin faster, by being in market when demand is real and dark when it is not. The seasonal-budget mistake is to cut spend when a vertical's cost per result rises in its off-season. The better move is to recognise that demand has shifted, not disappeared, and pivot the budget to where demand actually is.