NewWeather Demand Modelling is live. Forecast demand before it arrivesWeather Demand Modelling is live →
Pillar guide · Demand

The Weather Demand Modelling Guide

Weather is estimated to affect 25% to 35% of GDP in industrialised countries (Parnaudeau & Bertrand, 2018), yet most operators treat it as noise in the variance report. It is not noise. It is a measurable, forecastable demand signal you can read 30 to 45 days before it lands. This guide shows the maths, the published science, and the decisions it should change, framed for Australian seasons and the metric degree-day base.

Brendan Ellich
Brendan Ellich, Co-Founder & CEO
Blufire
Guide·16 min read·Updated June 2026Profit Velocity series
What this guide covers
  1. Why weather is a first-class signal
  2. The L.E.K. WISE framework
  3. Degree-days from an 18°C base
  4. Lag structure: HVAC vs roofing
  5. Pearson lag correlation method
  6. Why you must control for climate
  7. Vertical sensitivity table
  8. Forecast horizons and the 30-45 day window
  9. Proven accuracy uplift
  10. What it changes about your spend
01 · The premise

Weather is not noise. It is a demand signal.

When a CFO sees cost per acquisition jump 40% in a single month, the reflex is to call the account inefficient and cut budget. Often the account did nothing wrong. The weather moved the demand floor out from under it. Reading weather correctly turns a panic into a pivot.

The macro case is large. Weather conditions are estimated to affect 25% to 35% of GDP in industrialised countries, and at the retail level weather materially shifts store traffic, spend and basket size (Parnaudeau & Bertrand, Understanding the Economic Effects of Abnormal Weather, 2018; quantified for retail in Rose & Dolega, It's the Weather, Applied Spatial Analysis and Policy, 2021). The pattern holds in Australia: the Reserve Bank notes that weather and climate variability move farm output and aggregate GDP from year to year, with drought alone capable of trimming total GDP by around 0.15% (RBA, Climate Change and the Economy, 2019). The consumer-decision case is just as large: an oft-cited estimate attributed to The Weather Company and the US National Retail Federation puts roughly 62% of purchase decisions under some weather influence, across the four basic choices of what, where, when and how much to buy.

The point of this guide is not to repeat those headline numbers. It is to give you the method underneath them: how to turn temperature, precipitation and storm events into a forecast of your demand, how far ahead that forecast is honestly reliable, and what decision it should change. Throughout, we use only published, textbook methods. The maths here is degree-days, Pearson correlation, lagged regression and marketing-mix logic. There is no secret in the arithmetic; the edge is in applying it precisely and acting on it early.

The Blufire rule

High CPA correlated with weather does not mean cut budget. If cooling demand fell because temperatures dropped, the buyers who remain are still worth winning, and the right move is to pivot spend toward the in-season vertical (heating, in that example) rather than scale the whole account down to a shrinking demand floor.

02 · Framework

The L.E.K. WISE framework, and why we mirror it

L.E.K. Consulting's Weather Index for Service Essentials (WISE) is a consulting-grade, named framework for weather-driven services demand, and it validates the exact build a credible model needs (US framework; L.E.K. Consulting, A WISE Approach to Weather-Related Services Demand). It is a US method, but the structure transfers cleanly to Australian regions once you swap in local seasons, events and geography. It has three components:

  • Degree-days from an 18°C baseline (the metric base used by the Bureau of Meteorology; US sources state the equivalent 65°F). Cooling demand accrues above the base, heating demand below it. This converts raw temperature into an energy-and-service-relevant quantity.
  • Extreme-event tracking. Counts of hailstorms, floods, tropical cyclones, East Coast Lows and high-wind events, which drive same-period spikes in storm-response trades.
  • Regional granularity. Tropical northern Australia (QLD, NT, northern WA), temperate southern Australia (VIC, TAS, SA, ACT) and the eastern seaboard (NSW) behave very differently, so a national average hides the signal. Geography is the unit of analysis.

The build that follows from WISE is simple to state and hard to do well: regress demand on degree-days and event counts, at regional granularity, with the correct lag. Everything in this guide is an elaboration of that one sentence, applied to Australian conditions.

03 · The unit

Degree-days: turning temperature into demand

A degree-day measures how far, and for how long, the outside temperature sat away from a comfort baseline. The Bureau of Meteorology publishes Australian heating and cooling degree-day maps on an 18°C base, the point at which a typical building needs neither heating nor cooling (Bureau of Meteorology, Heating and cooling degree days; this is the metric degree-day base, equivalent to the 65°F used in US sources). Above it you accumulate cooling degree-days (CDD); below it, heating degree-days (HDD) . BOM's maps show cooling degree-days rising towards tropical northern Australia (Brisbane averages about 1,000 CDD a year) and heating degree-days rising towards the temperate south (Melbourne about 250 CDD, with far higher HDD).

The formula
Tavg = (Thigh + Tlow) / 2

CDD = max( Tavg − 18 , 0 )
HDD = max( 18 − Tavg , 0 )
The daily mean is the simple average of the day's high and low, in degrees Celsius. If the difference is negative, the degree-day count for that day is zero, not a negative number. Days accumulate across a season, which is why a hot January or February (summer) or a cold July (winter) compounds into a large demand signal rather than a single spike.
Worked example · one summer day (January)
1Record the day: high 33°C, low 21°C.
2Mean temperature: (33 + 21) / 2 = 27°C.
3Cooling degree-days: 27 − 18 = 9 CDD (positive, so it stands).
4Heating degree-days: 18 − 27 = −9, clamped to 0 HDD.
Result: that single day contributes 9 cooling degree-days. A week of similar summer days is ~63 CDD, the kind of accumulation that precedes a measurable lift in air-conditioning service calls.
Cumulative cooling degree-days across an Australian summerDemonstrative data
Illustrative accumulation of CDD from late spring (November) into the hottest months (January and February). The curve is convex: degree-days build slowly, then steepen through the peak of summer. Demand for cooling service tracks the accumulated figure, not any single day, which is why an early read of the season's trajectory is worth more than a daily weather check.

Two operational thresholds are worth committing to memory because they appear repeatedly in load and service data: cooling response tends to engage above roughly 29°C, and heating response below roughly 4°C (Yes Energy load-forecasting analysis; EngineerFix degree-day reference, converted to metric). Those are useful bid-modifier triggers, but they are blunt. The degree-day accumulation above is the quantity that actually scales demand.

04 · The mechanism

Lag structure: roofing spikes the same week, HVAC lags a quarter

The single most useful thing weather modelling tells you is not whether demand will move but when. Two flagship verticals show opposite timing, and confusing them is the most common modelling error in the category.

HVAC has a one-quarter lag. Degree-days accumulated in a given quarter lead to increased HVAC shipments and service the following quarter (L.E.K. WISE, US framework). There is also an immediate, hours-scale response in raw load: peer-reviewed load modelling finds outdoor temperature observed 1 hour and 24 hours earlier is the strongest predictor of demand, with a heating-degree-hour index reaching R² = 0.981 and a cooling-degree-hour index R² = 0.904 (PMC, predictive HCDH modelling study, 2023). So HVAC has a fast load response and a slow purchase-and-install response, and the marketing decision lives in the slow one. In the Australian calendar, cooling degree-days build through November and December, so the install peak lands across the height of summer (January and February); heating degree-days build through autumn, so the heating install peak lands in the depths of winter (June to August), concentrated in temperate southern Australia.

Roofing is a same-period spike followed by a larger decline. Extreme weather in a quarter raises roofing demand in the same quarter (storm-damage response), then triggers a larger-magnitude decline two quarters later as the storm pulled demand forward and exhausted the pipeline (L.E.K. WISE, US framework). In Australia the storm-damage trigger is seasonal: damaging hail in the eastern states (NSW, VIC) is most common from spring into early summer, peaking around November and December (Bureau of Meteorology; IAG hail data), while the tropical cyclone and severe-storm season across northern Australia runs November to April (Bureau of Meteorology). A major hailstorm can produce 4 to 8 times baseline lead volume in the days after the event (US Tech Automations storm-lead analysis, US data). The asymmetry matters: a roofing operator who scales spend on the post-storm boom and forgets the pull-forward will overspend into the trough.

Seasonal demand shape by vertical, Australian calendar (12-month index, Jan to Dec)Demonstrative data
Cooling / AC repairJan-Feb peak
Heating / furnaceJun-Aug peak
Roofing / stormNov-Dec peak
Cooling and heating service mirror each other across the Australian year: cooling peaks in summer (January and February), heating in winter (June to August). Roofing concentrates around the spring-to-early-summer storm season (November and December in eastern Australia) with a sharper, event-driven peak. Shapes are illustrative of the published patterns (L.E.K. WISE, US framework, applied to AU seasons; Bureau of Meteorology storm timing). The practical read: a single account that sells both heating and cooling never truly goes quiet, it rotates, which is exactly why pivoting beats cutting.
05 · The method

Pearson lag correlation, in plain maths

To make any of this defensible you need a number that says how strongly a weather feature moves demand, and at what delay. The textbook tool is the Pearson correlation coefficient computed at a series of lags, followed by a lagged regression. Nothing exotic; the rigour is in doing it at the right lag and controlling for the right things.

Pearson correlation at lag k
rk = Σ ( xt−k − x̄ )( yt − ȳ )  /  √[ Σ( xt−k − x̄ )² · Σ( yt − ȳ )² ]
x is the weather feature (say, degree-days), y is demand, and k is the lag in periods. You compute r at k = 0, 1, 2, … and the lag with the largest |r| tells you the response delay. r ranges from −1 to +1; values near zero mean no linear relationship at that lag.

What this yields in practice is instructive. In a peer-reviewed air-conditioning load model, Pearson coefficients ranged from strongly negative for indoor temperature setting (−0.94) and solar altitude (−0.71) to strongly positive for the load in the previous hour (+0.94) and previous 24 hours (+0.74) (Nature Scientific Reports, GRU/IASO AC-load model, 2025). The lesson hidden in those numbers: autoregressive terms dominate. Yesterday's demand predicts today's demand better than weather does. Weather is the exogenousdriver you layer on top of the demand's own momentum, not a standalone oracle. A model that ignores the autoregressive base will mis-attribute ordinary persistence to the weather.

Worked example · finding the lag
1Take weekly degree-days (x) and weekly booked installs (y) for an HVAC account over a full year.
2Compute r at lag 0 (same week), lag 1, lag 2 … up to lag 13 (one quarter).
3Suppose r peaks at r₁₃ = 0.61, so the strongest relationship sits roughly a quarter out, matching the WISE lag.
4Square it: R² = 0.37, so degree-days alone explain about 37% of the variance in installs at that lag, before adding autoregressive and event terms.
Read: a 0.61 correlation is meaningful but partial. It justifies launching campaigns a quarter ahead of accumulated degree-days, and it warns you that 60%+ of the variance is driven by something else (price, competition, the account's own momentum). Honest modelling reports both halves.

The published retail literature uses the same machinery at lower frequency: multiple linear regression with autoregressive terms (MLR-AR) and ordinary least squares to isolate weather coefficients, with snowfall in particular showing pronounced lagged effects in fashion retail, and one consumer-mood study finding weather optima roughly three days ahead of the behaviour it predicts (ResearchGate empirical daily-weather retail study, 2019).

06 · The trap

Control for climate, or you will over-claim

This is the failure that separates a credible weather model from a vanity one. Climate is the baseline seasonality of a place; weather is the deviation from it. If you regress demand on raw temperature without first netting out the normal seasonal pattern, ordinary summer-versus-winter swings get falsely attributed to weather sensitivity, and categories that are merely seasonal look weather-driven (ResearchGate / MDPI, Accounting for Climate When Determining the Impact of Weather on Retail Sales, 2023).

The fix is to model demand against the anomaly, the difference between observed conditions and the climatological normal for that week and region, rather than the raw reading. A 32°C day in Brisbane in January is climate; a 32°C day in Hobart in October is weather. Only the second should move your model.

Method note

The same caution explains why the published It's the Weather study (Wood et al., 2021) used random-forest modelling and reported the weather features lifting predictive R² from 0.8591 to 0.8942 on total sales: most of the explained variance was already in calendar and store structure, and weather added a real but modest increment on top. Anyone claiming weather explains the majority of demand is almost certainly failing to control for climate.

07 · The map

Which verticals move, and by how much

Not every business should build a weather model. The first question is whether your demand is weather-elastic enough to be worth the effort. The ranking below orders verticals by their peak demand lift under extreme conditions.

Peak demand lift under extreme weather, by vertical
HVAC / home svc80%Home & Garden65%Automotive55%Food & Beverage47%Travel & Tourism42%Retail Fashion37%
Peak lift in extreme conditions (US data; WeatherTrigger summary; L.E.K. WISE; US Tech Automations). HVAC and home services lead at +80%+; retail fashion sits at the bottom near +37%. These are peak figures, not annual averages. The vertical ranking transfers to Australia; the months in which each peak lands flip to the Southern-Hemisphere calendar.
VerticalPeak liftDominant signalTiming
HVAC / home services+80%+Degree-days, heat/cold extremes1-quarter lag (install)
Home & Garden+60-70%Temperature, precipitationSame season
Automotive+50-60%Storms, temperature swingsSame period
Food & Beverage+45-50%Heat, precipitationSame day
Travel & Tourism+40-45%Seasonal temperatureWeeks ahead
Retail Fashion+35-40%Temperature, snowfall (lagged)Lagged days
Roofing / siding4-8x per stormHail, high wind eventsSame week, then 2-qtr decline

Search demand is the cleanest measurable proxy a client can act on, because it is observable in real time and maps directly to campaign decisions. The peak-to-valley swings are dramatic in trades that respond to temperature extremes, and muted in trades that respond to discrete storm events.

Peak-to-valley search-demand variance by service keyword
Frozen pipe repair609%Heating system repair594%Emergency AC repair393%AC repair266%Circuit breaker219%Emergency plumber191%Furnace repair137%Emergency roof repair70%
Seasonal swing from trough to peak (US data; WebFX, Seasonal Search Shifts in Home Services Demand). Frozen-pipe and heating-repair queries swing +594% to +609% and peak in deep winter; emergency AC peaks in deep summer at +393%; roofing queries are far more muted, mostly under +70%. Read the months on the Australian calendar: heating and frozen-pipe demand peaks June to August, emergency AC peaks January and February. The implication: HVAC and plumbing budgets should breathe by 5-6x across the year, roofing budgets far less.
08 · The window

How far ahead can you honestly see?

A forecast is only as useful as it is reliable, and weather signals decay at very different rates. Pretending you can forecast a specific day's temperature three months out is how models lose credibility. Here is what each signal supports:

Credible lead time by signal type
Precipitation5dTemperature10dCampaign launch lead45dSeasonal pattern180d
Reliable horizon by signal (WeatherTrigger; WebFX). Precipitation is trustworthy ~5 days out, temperature ~10 days, and seasonal patterns ~3-6 months. The actionable campaign window sits between them: launch roughly 30-45 days before peak.

This is the heart of the practical recommendation. You cannot wait for the 10-day temperature forecast to launch a campaign, because creative, budget approval and the platform learning phase all take time, and the install itself lags. You launch on the seasonal pattern, which is reliable months out, and you tune with the short-range forecast as it arrives. On the Australian calendar that means cooling campaigns live by October and heating campaigns by April, roughly 30 to 45 days before demand begins to climb into summer and winter respectively.

The 30-45 day launch window, against the demand curveDemonstrative data
now
The solid line is observed demand to date; the dashed line and shaded band are the forecast and its widening uncertainty. The marked point is the launch decision: committed on the seasonal trajectory before the peak is visible, while the confidence band is still wide. Acting at the marker, not at the peak, is what converts a forecast into margin. Illustrative.
09 · The payoff

What adding weather actually buys you

The honest answer is: a real but bounded improvement in forecast accuracy, larger in some sectors than others. The published numbers are specific and worth quoting precisely rather than rounding up.

Forecast-accuracy uplift from adding weather data
Grocery20.2%Casual dining12.2%Home improvement3.3%
Accuracy improvement by sector (ToolsGroup). Grocery gains +20.2%, casual dining +12.2%, home improvement +3.3%. Aggregate forecast improvement averaged ~3.8% across five nations, rising to ~5.8% where local weather was most predictable; daily-sales weather effect reached 23.1% by store location and 40.7% by sales theme.

Notice the spread. A grocer gains an order of magnitude more from weather data than a home-improvement retailer, because the grocer's basket (ice cream, soup, salad, grilling supplies) flips with the daily forecast, while a roof gets replaced on a storm's schedule regardless of next week's temperature. The size of the prize is itself weather-sensitivity-dependent, which is why the sensitivity map above is the first thing to build, not the last.

Where this connects to Profit Velocity

Weather modelling raises Profit Velocity (the rate at which marketing and sales effort converts into durable contribution margin) by shrinking the denominator. Catching in-season demand early means winning buyers at a lower marginal cost, and not wasting spend defending a demand floor that has already shifted to another vertical. The gain is timing, not volume: the same dollars, deployed 30-45 days earlier and aimed at where demand actually is.

10 · The decision

What it should change about your spend

A weather model earns its place only if it changes a decision. Three changes recur:

  • Pivot, do not cut. When CPA rises with a weather signal, move budget to the in-season vertical rather than reducing total spend. The buyers who remain are still profitable to win.
  • Launch on the season, tune on the forecast. Commit campaigns 30-45 days before peak on the seasonal pattern, then adjust bids with the 5-10 day forecast as it firms up.
  • Respect the roofing pull-forward. Scale into a post-storm boom, but plan for the larger trough two quarters out so you are not still bidding aggressively into an exhausted pipeline.
i Heat Cool
Weather-led HVAC, real result
112%YoY growth

i Heat Cool. An Australian seasonal HVAC operator where demand rotates between heating and cooling across the year. Aligning spend to in-season demand rather than scaling down in shoulder months supported 112% YoY business growth and an A$180 cost per converted lead at an A$15k average job value. The principle is exactly the one in this guide: follow demand as it pivots, do not retreat to the floor.

None of the maths here is proprietary. Degree-days, Pearson correlation and lagged regression are textbook. The edge is in the discipline: control for climate so you do not over-claim, find the true lag so you launch at the right time, and act on the seasonal signal while the confidence band is still wide. That is what turns a weather report into a margin decision.

Primary sources
  1. Bureau of Meteorology, Heating and cooling degree days (18°C base, Australian climate averages). bom.gov.au
  2. Bureau of Meteorology, Australian tropical cyclone season (November to April) and severe-storm timing. bom.gov.au
  3. Reserve Bank of Australia, Climate Change and the Economy, 2019 (weather variability and Australian GDP). rba.gov.au
  4. L.E.K. Consulting, A WISE Approach to Weather-Related Services Demand (US framework). lek.com
  5. Rose, N. & Dolega, L., It's the Weather: Quantifying the Impact of Weather on Retail Sales, Applied Spatial Analysis and Policy, 2021. link.springer.com
  6. Parnaudeau, M. & Bertrand, J-L., Understanding the Economic Effects of Abnormal Weather (weather affects 25-35% of GDP in industrialised countries), 2018. researchgate.net
  7. U.S. Energy Information Administration, Degree-days (US 65°F base reference). eia.gov
  8. U.S. National Weather Service, Heating and Cooling Degree Days. weather.gov
  9. PMC, predictive heating/cooling-degree-hour modelling study, 2023. pmc.ncbi.nlm.nih.gov
  10. Nature Scientific Reports, ML AC-load demand prediction (GRU/IASO), Pearson sensitivity, 2025. nature.com
  11. MDPI / ResearchGate, Accounting for Climate When Determining the Impact of Weather on Retail Sales, 2023. mdpi.com
  12. ResearchGate, The impact of daily weather on retail sales: an empirical study in brick-and-mortar stores, 2019. researchgate.net
  13. ToolsGroup, Using Weather and Climate Data to Improve Demand Forecasting. toolsgroup.com
  14. WebFX, Seasonal Search Shifts in Home Services Demand (US data; read the months on the Australian calendar). webfx.com
  15. US Tech Automations, storm-lead analysis, 2026 (US data). ustechautomations.com
  16. Yes Energy, electrical load forecasting / AC temperature sensitivity. yesenergy.com
Related guides and reports
Guide
The Contribution Margin Playbook

Revenue tells you a business is busy. Contribution margin tells you whether it is worth running. The operator's guide to CM1, CM2 and CM3, the payback maths behind every acquisition dollar, and the one metric that ties marketing effort to durable profit.

18 min read · June 2026 →
Guide
The Profit-Led Measurement Guide

Your ad platforms collectively claim more sales than your business actually made. This guide is the operator's manual for measuring what your marketing truly causes, denominated in profit rather than reported revenue, using the four standard methods and the maths that ties them together.

14 min read · January 2025 →
Guide
The Retention & LTV Guide

Most ecommerce customers buy once and disappear. The median DTC repeat rate is 18.8%. This guide is the operator's manual for the other 20%: how retention curves behave, how to fit them, what a second order is worth, and how owned channels turn retained customers into acquisition you have already paid for.

15 min read · May 2026 →