Methods, sources and assumptions

Every input below is labelled sourced (published or measured data, with the source), modelled (computed from sourced data by the method described) or assumed (a judgement, stated with its value and reason so it can be challenged and, as check-ins accumulate, refitted). Uncertainty is given as the 25th–75th percentile of Monte Carlo draws wherever the inputs allow.

1. The Park Day Index

The day (0–10) is a weighted sum of factor scores, each 0–10, then capped by the most limiting factor:

Weather carries 84% of the weight. Limiting factors assumed, following the override rules of the Climate Index for Tourism (de Freitas, Scott & McBoyle 2008): each factor score maps to a highest allowed index and the most limiting one wins. Temperature 0→2, 3→4.9, 5→6.5, 7→8, 8.5→10; rain 0→1.5, 3→3.5, 6→5.5, 8→7.5, 9.5→10; wind 0→2, 3.5→6, 7→10; grey sky (raw sunshine) 0→6.2, 1.5→6.45, 4→10; wet ground 0→5, 3→6.5, 5→10.

Calibration. The weighted score is mapped onto 0–10 by piecewise-linear quantile mapping on 1,188 real days (July 2023 – September 2026) at the park, knots 1.30→0, 4.30→3.5, 5.83→5, 7.23→6.5, 8.63→8, 9.97→10. With the limits applied, those days fall 24% Stay home, 15% Marginal, 28% Fair, 19% Good and 14% Excellent. Bands: Stay home under 3.5, Marginal 3.5–5, Fair 5–6.5, Good 6.5–8, Excellent 8 and over.

Helping and hurting. For each factor, today's index minus the index with that factor at its 2023–26 average. "Limiting factor" appears when a cap applies.

The headline range. The number is the central forecast. The ten-square bar is solid to the 25th percentile and hatched to the 75th percentile of 1,000 draws over the 51 ensemble runs and the party model's uncertain inputs. Confidence (in the stories): the share of the 51 ensemble runs that put the day within a point of the forecast, every run scored the same way: high at 70% or more, medium at 40–70%, low below 40%.

2. Weather data sourced

• Forecast: ECMWF IFS, the European Centre for Medium-Range Weather Forecasts' high-resolution global model on its native ~9 km grid, hourly, via Open-Meteo for −37.789, 144.981. At this point Open-Meteo's "best match" is the ECMWF IFS. It is the most skilful openly available global model; the Bureau of Meteorology's own models are not available through an open API for this location (ACCESS-G was tested and returns no data). New runs arrive about 7 hours after their 00 and 12 UTC start times.
• Ensemble: ECMWF IFS ensemble (ENS, 0.25°), 51 runs from slightly different starting conditions, about 15 days, via the Open-Meteo Ensemble API. Their spread is the forecast's uncertainty.
• Sunlight: hourly direct normal and diffuse irradiance from the same forecast, used for the sun on a person (section 6).
• Air quality: Copernicus Atmosphere Monitoring Service (CAMS) PM2.5 forecast via Open-Meteo, while it forecasts (about four days).
• Past weather: hourly rain and FAO-56 reference evapotranspiration for the previous 60 days; ERA5 reanalysis (Hersbach et al. 2020) for normals.
• For comparison: Bureau of Meteorology Melbourne forecast.

3. The park: trees, buildings, terrain

4. Sun and shade modelled

Sun and shade do not depend on the weather, so they are computed in advance: every night, for each of the coming eight days. Sun position comes from a standard solar ephemeris every 4 minutes from sunrise to sunset, and more often at low sun so that no shadow moves more than 4 m between samples.

For each sample:

• Trees: every crown slice and trunk is projected along the sun's ray onto a 50 cm grid, allowing for sloping ground, and attenuates the beam by Beer–Lambert extinction with the tree's own leaf angle (section 5).
• Buildings cast shadows by ray-marching their heights across the whole surrounding area, not only the park.
• Terrain shades itself; the slope of the ground changes how much beam a cell receives. Every spot is judged from its own ground: a building's shadow is measured from its roof down to the receiving spot's elevation, and tree shadows run further onto lower ground. The raised bank at the north-east corner of the W. T. Peterson Oval, about 4 m above the oval, therefore sees over roofs and trees that shade the oval below it.
• Far horizon: trees and buildings from 400 m to 3 km enter as a horizon angle for each 2° of azimuth, so the evening sun can set behind distant roofs and trees. The angles are computed every 25 m across the park and then corrected for each spot's height above those sample points, using the distance to the obstruction that sets the angle: 4 m of height lowers a roofline 400 m away by about 0.6°, around three minutes of evening sun.
• Penumbra: the sun is a disc 0.533° across, so the edge of a shadow cast from h metres up is blurred over h·0.0093/cos z across the shadow and /cos² z along it. Cells partly covered mix light, T = (1 − w) + w·e^−τ, not optical depth (which would let a sliver of dense crown black out a whole cell).

The results are each spot's direct-beam transmission through the day (1 m grid, used by the spot map), and day maps computed at 25–50 cm: sun energy, hours in shade, and the last minute of direct sun, interpolated between samples from the movement of the shadow edge. The ray tracing is run on a park-sized receiving area with every shadow source in range; this gives the same results as tracing the whole surrounding area, to within a second of last sun.

Resolution on the website. The website serves the day maps at 1 m. Each 2 × 2 block of computed cells becomes one cell: the mean for sun energy and hours in shade, and for last sun the lower median of the block's own times, so the time shown is one that occurs at that spot, not an average that occurs nowhere. The website samples these maps on a 2 m grid, so 1 m keeps twice the resolution it uses. The stories use the maps at the computed resolution.

Last sun. "Last sun" for the park is the 99.5th percentile of the open lawn's last minute of direct sun: the time by which all but 0.5% of the lawn has lost the sun, so a single stray cell does not set it. It is shown to the minute, rounded up. Late in the day several spots can tie, because they stay lit until the sun is half a degree above the horizon, where the model stops (below that the horizon is not reliable). The spots that keep the sun longest are marked: every spot within a minute and a half of the latest, up to three, at least 60 m apart. On an overcast or wet evening no last sun is given.

Spot ray trace. At a dropped pin the browser re-traces the sun every 5 minutes through each crown as a true ellipsoid and against a 1 m raster of building and ground heights.

5. Canopy optics and the seasons

Beam transmission through foliage: T = exp(−G(θ)/cos θ · Ω · L), with clumping Ω = 0.8 assumed and leaf area index L from the tree's leaf habit and the season assumed.

Leaf angle sets G(θ). Campbell's (1986) ellipsoidal distribution gives K = √(χ² + tan²θ) / (χ + 1.774(χ + 1.182)^−0.733), with χ from the mean leaf angle (Campbell 1990): MLA = 9.65(3 + χ)^−1.65 radians.

Trees without a species record keep the spherical default (χ = 1).

The seasons. Deciduous leaf area follows a seasonal curve: full leaf November–March, falling April–June, bare in July, leafing out September–October assumed, with LAI = 0.5 + leaf fraction × (summer LAI − 0.5) for the bare branches. Every shade calculation uses the leaf area for the date shown.

6. Feels-like and comfort

In the shade: the forecast's apparent temperature, the Bureau of Meteorology's measure (Steadman 1994): AT = Ta + 0.348e − 0.70v − 4.25, with Ta the air temperature (°C), e the water-vapour pressure (hPa) and v the 10 m wind (m/s). Near skin temperature wind stops cooling, so the wind term fades out between 25 and 35 °C: a hot northerly is no relief.

In the sun sourced method: Steadman's full form adds 0.70·Q/(v + 10) for Q W/m² of sunlight absorbed per square metre of body. Q is the direct beam falling on a seated person's sunlit silhouette, Q = 0.7 × f_p × DNI, with clothed-body absorptivity 0.7 and the projected-area factor of a seated person f_p falling from 0.26 with the sun low to 0.20 overhead (Fanger 1970; ISO 7726). DNI is the forecast direct normal irradiance. On a calm clear noon this adds about 8 °C; in a 25 km/h wind about 5 °C; with the sun 5° up at a clear sunset about 2 °C. At a spot, the beam is multiplied by that spot's transmission, and a spot sheltered from the wind gets back the wind chill it is spared.

Comfort assumed: 10 from 22 to 26 °C feels-like (the preferred range for people sitting outdoors in light clothing); 0.3 points less per degree from 22 down to 18 °C (cool enough to want the sun); 0.45 per degree from 26 to 32 °C, the UTCI's band of moderate heat stress (Bröde et al. 2012), in which people still sit out; then 1 per degree above 32 °C and 1.4 per degree below 18 °C. Cold costs more because people sit on the ground. The day's comfort uses the warmest sunlit feels-like, approximated as air + 2.5 °C × (1 − cloud fraction) − 0.15 °C per km/h of wind (faded near skin temperature).

7. Wind at the blanket modelled

The forecast 10 m wind is brought down to 1 m with a neutral log profile over grass, u(1)/u(10) = ln(1/z₀)/ln(10/z₀), z₀ = 0.03 m (60%), then multiplied by the spot's shelter for that wind direction.

Shelter. Buildings (solid) and crowns (porous, optical porosity e^−0.35·LAI + 0.1 assumed) within 300 m upwind act as windbreaks. Relative wind behind a barrier of height H follows a parametric fit to the ranges summarised by Heisler & DeWalle (1988): u_min = 0.15 + 0.6φ at x_min = (2 + 6φ)H, back to 0.8 at x_80 = (10 + 20φ)H, with a small dip within about 3H upwind. Each spot takes the strongest shelter along five parallel upwind rays (0, ±6, ±12 m) and averages them; a single tree's wake is weakened to 60% assumed because it is a point obstacle, not a long windbreak. Computed for 16 directions on a 4 m grid. Not modelled: flow deflection, channelling, gusts and turbulence.

8. Hour by hour and the best time

Each hour from 6 am to 9 pm is scored on sunshine 19%, comfort 24% (the better of sun or shade), rain 20%, wind 15%, light 10%, wet ground 5%, mosquitoes 4% and parties 3% assumed, through the same calibration and the same limiting factors as the day. Rain falling in the hour (0.1 mm or more in the central forecast, with a stated chance of 40% or more, or 0.2 mm or more) counts as at least a 90% chance; hour caps (rain score → highest hour score): 0→1, 5→3, 7→6.5, 9→10, so a 50% chance or rain falling is Stay home. Wet ground caps an hour: 0→4.5, 3→6, 5→10.

The best time is one rule, on the website and in the stories alike. The hours from first sun (to the nearest hour) to the last full hour of sun are put into their bands (Stay home, Marginal, Fair, Good, Excellent), and a run shorter than two hours is folded into a neighbouring run one band away. The best time is the whole run that holds the day's best hour (if several hours tie, the middle one). It comes from the same hourly readings as the one-line summary and the High and Low, so the three always agree; it reads "all day" when the run spans the whole window. If even the best hour is under 5, there is no good time: the stories say so, and the website names the least-bad stretch. The Golden hour persona keeps its own window, golden hour itself.

The one-line summary names the bands through the day ("Good until 11 am, stay home after"; runs under two hours are folded into a neighbour one band away), stops once the day turns to Stay home, then adds the weather in the Bureau's words: light winds, breezy (15 km/h+), gusty (gusts 45 km/h+), blustery (gusts 45+ and at least 1.8 × the mean), windy (28+), very windy (40+), damaging gusts (65+); a cool change when the wind swings from the north into the south-west within three hours, at least 10 km/h, with 2.5 °C or more of cooling; storms and showers by start time.

9. Spot by spot: the map modelled

The map scores the lawn each hour for the persona shown: every square metre in the stories, and 2 m squares on the website (2.5 m on smaller screens). It starts from the park's hour score and replaces what differs from place to place:

• Sun: the spot's own direct-beam transmission (section 4), in its feels-like (section 6) and in its sunshine. A shaded spot misses the beam's share of the light reaching a seated person, computed from the forecast beam and diffuse light: about 60% at a clear noon, nearly 90% at a clear sunset.
• Wind: the spot's shelter (section 7).
• Light: with the sun low, a spot still in the sun is in full light and one already in shade is in the park's dimmer light, in proportion to the beam reaching it.
• Noise: roads, trams, paths, the skate park, the playgrounds, the courts, the bowls club and events at the rotunda that hour (section 10); 0.12 points per dB(A) above 55, at most 2.5, because people sitting in a park are there partly for quiet assumed.

Then each spot loses points for what is near it assumed:

Paths, the ovals, courts, bowling greens, playgrounds, the raingarden, car parks and the buildings in the park are not scored: nobody sits there.

The isobars. The scores are smoothed with a 1.5 m Gaussian over the lawn only (about a picnic blanket), then contoured every half point and filled with the index colours. H marks the broad area of highest scores; L marks a broad low inside the park only when it is below 5. When the stories reach the last sun, the index fades and the day's last-sun map takes its place: filled bands at 20, 15, 10, 7, 5, 4, 3, 2 and 1 minutes before the last sun, darker for later (smoothed over 1.5 m; patches under 40 m² left out), each line labelled with the time the sun lasts until. The map's symbols mark the last-sun spots, swooping hotspots, likely party set-ups and, at dusk, mosquito habitat.

Grid references are the site's 20 m squares, lettered west to east from A and numbered north to south from 1.

10. Noise modelled

Roads: Calculation of Road Traffic Noise (Department of Transport 1988). L10(18h) = 29.1 + 10 log₁₀ Q, speed and heavy-vehicle correction 33 log₁₀(V + 40 + 500/V) + 10 log₁₀(1 + 5p/V) − 68.8, distance −10 log₁₀(d′/13.5), angle of view 10 log₁₀(θ/180), −10 dB behind a building assumed; LAeq ≈ L10 − 3. Traffic: Department of Transport and Planning AADT 2019 sourced; uncounted local streets 1,500 vehicles a day assumed; a typical urban hourly profile assumed.

Trams: pass-bys at 80 dB(A) SEL at 7.5 m, 16 an hour assumed; LAeq,1h = SEL + 10 log₁₀ N − 35.6. Paths: 47–50 dB(A) at 3 m on a busy afternoon assumed.

The skate park: 64 dB(A) LAeq at 10 m when in use, falling 6 dB per doubling of distance; busiest after school and on weekend afternoons; silent in rain assumed, in line with published skate park noise assessments (about 60–65 dB(A) at 10–15 m). The playgrounds: 60 dB(A) LAeq at 10 m when busy, the same fall-off; busiest after school and through weekend days; empty in rain assumed, in line with published playground and schoolyard measurements (about 55–65 dB(A) at 10 m). The rotunda: an event (a band, a ceremony, an amplified speaker) at 72 dB(A) LAeq at 10 m; the hour's expected level adds 10 log₁₀ of the chance one is on (section 11) assumed. The courts and the bowls club: the basketball court 60 dB(A), the tennis club 55 dB(A) and the bowls club's social bowls and bar 60 dB(A) LAeq at 10 m when busy, the same fall-off, busiest after school, after work and on weekends, silent in rain assumed, in line with published outdoor court assessments. Ground absorption and reflections are ignored.

11. People and park life

Children's birthday parties modelled

expected parties = Σ suburbs (children aged 1–9 × share choosing Edi Gardens) × birthdays mapped to the day × share with a party × share held in a park × weather

• Children: ABS 2021 Census G04 single year of age, 15 inner-north suburbs, 12,859 children sourced.
• Birthdays: ABS births by month of occurrence, Victoria 2016–2024 sourced.
• 85% of parties move to the nearest weekend, split 55/45 Saturday/Sunday; 60% have a party, 25% of those in a park, and 45/15/5% choose Edi Gardens by distance tier; rain, cold and heat reduce them assumed.
• Chance of at least one party = 1 − e^−expected (Poisson). Set-up within 35 m of playgrounds, barbecues and taps, audible to 80 m assumed.

Ovals and stray balls

W. T. Peterson Oval hosts Fitzroy Football Club, Fitzroy Junior Football Club and Edinburgh Cricket Club sourced. Fixtures are not open data, so seasons and typical slots stand in: cricket October–March, football April–September; Saturday matches, Sunday juniors, midweek training, Friday-evening junior cricket at Alfred Crescent Oval; 50% home games assumed. Stray balls: cricket 1.6 an hour of play with a mean carry of 12 m past the boundary, football 6 an hour and 7 m, spread around the perimeter assumed.

Dogs

City of Yarra rules: about half the park off-leash; no dogs on W. T. Peterson Oval; off-leash on unfenced sports grounds only when not in use; on-leash within 10 m of playgrounds sourced. Dogs per hour through the whole park: about 45 in the busiest dog hour on a dry, mild day, derived from the 6,791 dogs registered in the City of Yarra (Domestic Animal Management Plan 2021–25): about a quarter live within an easy walk, about 60% of those go to a park each day and 40% of them here, about 410 visits a day spread by hour, following weekday and weekend activity profiles, × (1 − 0.6 × chance of rain), and fewer on hot afternoons (6% less per degree above 30 °C) assumed, to be refitted from check-ins.

Mosquitoes

Index = season × time of day × temperature × wind × habitat × recent rain assumed, following known drivers: dusk and dawn activity, little activity below about 12–15 °C, flight suppressed by wind, breeding after rain with a one-to-two-week lag, November–April. Habitat is the raingarden and dense canopy. An index, not a trap count.

Swooping

Magpie breeding August–November, peak mid-September to mid-October; risk grows with nearby trees 14 m and taller (eucalypts counted double), with hotspots at least 80 m apart. Masked lapwings on open turf July–December assumed.

Hay fever

Grass pollen: a seasonal Melbourne curve, multiplied up on hot northerly days and down after rain, the drivers Melbourne Pollen uses; categories from Melbourne Pollen assumed curve, sourced categories. Hay fever follows grass pollen; falling elm seeds are noted alongside but are not an allergen.

Leaves and seeds

Autumn leaf litter is the fall in deciduous leaf area over the past 21 days (e-folding 10 days); elm seeds fall from early September, peaking mid-September to early October assumed; gusts over 30 km/h shake more loose. Index = 0.6 × litter × (1 + 0.5 × gust) + 0.4 × seeds × (0.7 + 0.6 × gust), 0–1.

Busyness assumed + modelled

There is no people counter in Edi Gardens, so this is an estimate to be refitted from check-ins. Expected groups on the grass in an hour = 200 × profile(hour) × e^{0.45 × (hour score − 6.5)} × an after-dark fade (e-folding 45 minutes after sunset) × e^{−4 × chance of rain} × up to −80% as the ground goes from damp to soaked. Weekdays have a lunch bump and an after-work rise; weekends build to a mid-afternoon peak. It is shown as a scale, not a headcount: against a typical fine weekend afternoon, Quiet (under a quarter of it), Moderate (to three quarters), Busy (to twice) and Very busy (beyond).

Where people sit is a seating game on a 4 m grid: groups arrive one at a time and choose a spot with probability ∝ exp(utility / 1.4), a logit rule (McKelvey & Palfrey 1995). Utility = spot quality − crowding + company. Quality: sun or shade as the hour favours; prospect and refuge, a tree at the back and a view over open grass (Appleton 1975; Gehl 1987), up to 0.7, peaking 1.5 m outside a crown's edge; up to 0.3 near (not on) a path; slope, noise and wind costs; the shops after 4 pm. Crowding: personal space as a Gaussian, σ = 4 m (Hall 1966). Company: up to +0.6 from groups 10–25 m away (Whyte 1980). As the lawn fills the best spots go first and spacing takes over (Fretwell & Lucas 1970).

The rotunda assumed

Fitzroy Memorial Rotunda (1925), the park's main place for performances, ceremonies and organised groups. Events arrive as a Poisson process: 0.012 an hour on weekdays, 0.035 on weekends and 0.05 on public holidays, scaled by an hourly profile, × 1.6 when 22 °C and over, × 0.5 under 14 °C, and reduced by the chance of rain. In 3D it is modelled from its heritage description (Victorian Heritage Database place 103220) at its LiDAR footprint (6.1 m across) and height (9.36 m).

The beer test modelled

Minutes for a 375 ml can from the fridge (4 °C) to reach 10 °C on a rug: a heat balance stepped every 10 seconds. Convection combines free and forced flow over a cylinder (Churchill–Usagi blend; Hilpert correlation for the wind at 10 cm, 0.21 × the 10 m wind); long-wave exchange with emissivity 0.8; absorbed sunshine (absorptance 0.5) on the can's silhouette from the beam, the sky and the ground (albedo 0.2), or 40% of the sky in shade; and the latent heat of dew condensing on a can below the dew point (Chilton–Colburn analogy). Heat capacity 4.04 kJ/(L·K) assumed; the base is taken as insulated. Shown in the sun when there is sun, otherwise in shade. 30 °C in the sun ≈ 9 minutes; 22 °C ≈ 15.

Shops

Groceries, delis, bakeries, bottle shops and takeaway within 900 m from OpenStreetMap sourced. Walking time: straight-line × 1.3 on the street grid to the park's edge, then the park's own paths.

12. Ground and grass modelled

Ground moisture (is the grass damp?): a 20 mm soil bucket per 50 cm cell over the past seven days. Rain in (30% intercepted under full canopy), evapotranspiration out scaled by that cell's sun, started from modelled topsoil moisture assumed parameters. During the day, wetness rises by 0.06 per mm of rain and falls by up to 0.015 per sunny hour.

Grass condition (is it green?): FAO-56 single-bucket water balance (Allen et al. 1998) over 60 days: root-zone available water 60 mm, depletion fraction 0.5, crop coefficient 0.8 for turf assumed; green until half the water is used, then browning to straw. The ovals are irrigated.

Mowing and sprinklers. The raingarden (2010–11) treats stormwater from the North Fitzroy Main Drain into underground storage that irrigates the park's trees; a 2020–21 upgrade raised capture from 4 to 24 ML sourced (City of Yarra 2021). Neither sprinkler times nor mowing schedules are published: sprinklers are assumed 3–6 am on irrigation-season nights without 5 mm of rain in the previous three days, inside Victoria's Permanent Water Saving Rules window; mowing has an assumed seasonal cycle (10 working days in spring) assumed.

13. Health and safety

• Heat health sourced: the Victorian Department of Health threshold for the Central district is a mean of 30 °C (the day's forecast maximum and the following night's minimum). Official alerts follow the Bureau of Meteorology's heatwave service, so this one is indicative.
• Sun protection times sourced: when the forecast UV Index is 3 or more (SunSmart, ARPANSA). Time to sunburn: 1 UVI = 0.025 W/m² of erythemally weighted irradiance (WHO 2002); one Standard Erythema Dose (SED) = 100 J/m²; minimal erythema doses of 2.0, 2.5, 3.5 and 4.5 SED for skin types I–IV assumed; in shade half the UV still arrives from the sky, UVI_here = UVI × (0.5 + 0.5T) assumed.
• Air quality sourced categories, modelled values: EPA Victoria hourly PM2.5 categories (good under 25, fair 25–50, poor 50–100, very poor 100–300, extremely poor 300+ µg/m³) applied to the CAMS forecast. These are not EPA measurements; see EPA AirWatch.
• Thunderstorms sourced: forecast WMO weather codes 95–99.
• Thunderstorm asthma modelled pointer, 1 October – 31 December: flagged when grass pollen is high or extreme and a thunderstorm or gusts of 60 km/h are forecast. Always defer to the Department of Health's epidemic thunderstorm asthma forecast on VicEmergency.
• Cool change: as defined in section 8, flagged in the two hours before it arrives.
• Fire danger and Total Fire Bans sourced: the CFA Central district feed, read daily.
• Falling branches assumed: on hot (30 °C+), still (mean wind under 12 km/h) afternoons, the summer branch drop pattern reported for large eucalypts, elms and planes, or when gusts reach 60 km/h.
• Park lighting: the park's lamps are currently out of order, so paths are dark after sunset.
• Alcohol sourced: City of Yarra local law, no drinking in parks 9 pm – 9 am, plus New Year's Eve restrictions.
• VicEmergency: the State's public incident feed within 15 km plus statewide warnings, refreshed every two minutes; planned burns are left out. A pointer only: follow VicEmergency and the emergency services.

14. Personas assumed

The Park Day Index is for everyone. A persona re-scores each hour with its own preferences and times, then finds its own best time and best spots:

• Everyone: sun or shade, whichever is better.
• Sun chaser: wants direct sun and will put up with a cool day for it.
• Shade seeker: sits in shade; cloud is fine; heat is the enemy.
• Golden hour: here for the last light and the sunset.
• After work: weekdays from 5:30 pm, any time on weekends and public holidays.
• Early bird: before the crowds, 6 to 10 am.
• Picnic dinner: weekday evenings from 5:30 pm, weekends from 4 pm, near the shops.
• Rain lover: rain and wet grass don't count against an hour.
• My persona: set sun or shade, heat and cold tolerance (±3 °C), wind sensitivity, children's parties, quiet, toilets, golden hour and free hours. Stored only on your device.

15. The week, the outlook and normals

The week. Each of the next seven days is scored exactly like today, from its own hourly forecast, party model, pollen, swooping season and a ground bucket projected forward. The number is the median across the 51 ensemble runs, with the 25th–75th percentile; beyond the ensemble, a single run with the range widened by 0.35 points per day beyond day 3 assumed.

Days 8–16: the deterministic forecast, with the ensemble range where it reaches. Beyond the forecast: sun times are exact; the index, temperature, rain, golden-hour odds and pollen are the 1991–2020 normal for the date.

Normals: 1991–2020, the WMO climate normal period, at this point from ERA5 reanalysis (Hersbach et al. 2020): 262,992 hours. For each day of the year a 15-day window across 30 years gives the 10th, 25th, 50th, 75th and 90th percentiles of temperature, feels-like, wind, gusts, rain, cloud, sun energy, pollen, party chance, and the Park Day Index itself (every one of the 10,958 days scored with the same code). ERA5 is a ~25 km grid, so its extremes are slightly muted.

Trend: how each day's index has moved as it got closer, from Open-Meteo's Previous Runs API and the Bureau's own forecast log. Every earlier forecast is re-scored with today's model, so the trend shows the weather forecast changing, not the model. ▲▼ show the change since the previous evening's forecast.

Public holidays (Nager.Date) behave as Sundays for parties, sport, mowing and traffic assumed.

Sun times: sunrise, solar noon, golden hour (sun from +6° to −4°) and sunset from a solar ephemeris accurate to about a minute (Meeus 1998), with Melbourne daylight saving. Golden hour rating assumed: sunset sky = 5 + 3 × bell(high cloud, 45%, 30) + 2 × bell(mid cloud, 35%, 25) − 5 × (low cloud above 25%, full by 75%) − 2.5 × overcast − 4 × chance of rain, 0–10; rating = 0.6 × sky + 0.25 × comfort + 0.15 × wind, capped by the same limits.

Moon and astrology: moon phase, rise and set from a low-precision lunar ephemeris, checked against 2026's new moon (10 October) and full moon (26 October). Sun and Moon signs in the tropical zodiac from their ecliptic longitudes; Mercury's apparent retrograde motion from its geocentric longitude (JPL approximate Keplerian elements), reproducing all three 2026 retrograde periods within a day. Astrology's weight in every score: 0%.

16. Uncertainty modelled

• Park Day Index: 1,000 draws over the ensemble runs and the party inputs.
• Children's parties: 2,000 draws; assumed shares drawn within plausible ranges (party 45–75%, park 15–35%, weekend shift 75–92%, Saturday 45–65%, each distance tier × lognormal σ = 0.35), weather from a random ensemble run.
• Best times: the share of ensemble runs that make each hour Good (6.5+).
• Sun until at a pin: 40 re-traces with crown radii × (1 ± 15%) and leaf area × (1 ± 20%) (1σ).
• Busyness: 400 draws an hour over the ensemble's hour scores and a lognormal size uncertainty (σ = 0.35).

Reported as the median with the 25th–75th percentile. Draws use a fixed seed, so the same inputs give the same numbers.

17. Testing, validation and accuracy

Sanity tests. Before anything is published, the scoring code is run against 92 checks a reasonable person would agree on, and the spot map against 41 more. Publishing stops if any fails. A selection, with today's results:

Also checked, among others: more rain, heat (above 30 °C) or cold (below 15 °C) never improves a day across 2,000 random days; cloud never makes a 35 °C day worse; a 95% chance of a party costs a perfect day at most 0.6; every real 38 °C+, gale and 10 mm+ day rates Marginal or worse and every mild, calm, dry, clear day Good or better. On the map: nobody is scored on a path, oval, court, playground or building; the street edge, the paths, the toilets (and downwind of them), the skate park, the playgrounds and the ovals in use all score lower nearby; on a hot afternoon shade wins and on a cool clear evening the last of the sun wins; each persona's best spots fall where its preferences say.

Against real crowds. Every component was recomputed for 1,188 days from the hourly archive at the park and compared with City of Melbourne's automated pedestrian counters in two parks (Birrarung Marr and Boyd Park), 10 am–6 pm, as "busier or quieter than usual for that weekday and month" (712 days after removing festivals, closures and faults). Park foot traffic against usual, by band: Stay home −15.0%, Marginal −3.6%, Fair +2.8%, Good +4.4%, Excellent +10.4%. Weather explains about 9% of the day-to-day variation in these counts; the counters are in city parks that are partly thoroughfares, so they understate how much weather matters to people who come to sit. A 500-draw Monte Carlo over the weights kept the rank order of days at Spearman 0.98: conclusions do not hinge on the exact weights.

Sun and shade. The fast ray-tracing path is checked against the full-area reference: identical beam transmission, last sun within a second, and identical best spots, including at low sun where shadows are longest.

Geometry check against aerial imagery and council surveys, 15 October 2026

Everything on the map was compared with independent data: Esri World Imagery at 0.2 m in the map's own projection (a 610 m square over the park), the City of Yarra street and park tree and significant tree registers (CC BY 4.0), and the ray-traced sun and shade for 18 October 2026.

18. Check-ins and accuracy

A check-in records the spot (grid reference), the minute, multiple-choice answers (sun, wind, feel, ground, crowd, child's party within earshot, dog visit, mosquitoes, noise, overall vibes) and the forecast for that spot and minute, frozen at submission so it can be scored. Check-ins are anonymous: a random per-device code, no names, no IP address, no position finer than a 6.7 m square. One-tap emoji check-ins accept only a fixed palette of 30 emojis, so there is no free text.

Accuracy is reported as exact agreement and agreement within one category, next to the baseline of always guessing the most common answer. Probability forecasts (a party within earshot, a dog visit, a rotunda event) are scored with the Brier score against the same baseline. As check-ins accumulate, the assumed rates (rotunda, parties, dogs, busyness) are updated by Bayesian refitting, and each check-in records the model version that made its forecast so changes can be scored before and after.

Vibes model. A learned correction is layered on the physics and used only when it measurably does better: ridge regression (Hoerl & Kennard 1970) on the difference between reported vibes (1–5) and the forecast, from the forecast features frozen in each check-in, with the ridge strength chosen by leave-one-out cross-validation; minimum 8 rated check-ins. The optional pulse survey's pairwise choices between two real spots fit a Bradley–Terry model (1952) once there are 10, estimating how people trade sun against noise, party risk and toilet distance. Learned numbers are shown with their sample size.

19. Access

Toilets sourced: National Public Toilet Map, for each facility, its hours and its access. The map's coordinates are geocoded and can miss the building, so a point with no structure within 10 m is placed on the OpenStreetMap toilet block within 50 m, where there is one assumed; checked against aerial imagery (section 17). "Edinburgh Gardens North": unisex, wheelchair accessible (left-hand transfer), baby change, drinking water; placed on its block, about 34 m from the map point. "Edinburgh Gardens South" (Alfred Crescent): unisex, ambulant, baby change; its map point is 78 m from any structure; checked on the ground, it is the block beside the pavilion, directly south of the basketball court, and is placed there. Both 6 am–9 pm October–March and 6 am–6 pm April–September. Nearest Changing Places: Bargoonga Nganjin North Fitzroy Library, 182–186 St Georges Road. Walking time to a toilet open at that minute: fastest route on a 2 m grid (grass 1.25 m/s, paths 1.4 m/s) modelled.

Step-free routes modelled: fastest route on a 2 m grid for a wheelchair, pram or walking frame: sealed paths 1.0 m/s, unsealed or unrecorded paths 0.6 m/s, grass 0.35 m/s, grass steeper than 1:14 (the maximum ramp gradient in AS 1428.1) at half that assumed. Slope from the 1 m LiDAR terrain; 1:20 or gentler counts as level assumed. Seats (36, 25 in the park, backrests recorded for 16) and drinking water from OpenStreetMap.

Accessible spot finder: grass in the park, not on a path or oval, gentler than 1:20, 1.5–15 m from a sealed path; ranked by path distance, nearest seat, step-free time to an open accessible toilet and noise.

20. Scales

21. Known limitations

1. Tree canopy is under-represented. The modelled canopy covers about 17% of the park; Vicmap Vegetation Cover 2018 measures about 35%, and aerial imagery agrees. The tree data comes from the 2017 LiDAR capture, which misses later plantings and merged crowns. The model therefore shows more sun and less shade than the park has, mostly in the north half and along the west edge. Better tree data (a later LiDAR capture or the council's park tree inventory) is being sought.
2. A row of trees on the park's west edge is missing from the tree data and has no sourced height. It lies on the line to the setting sun from the bank at the north-east corner of the W. T. Peterson Oval, so last sun there carries about ±2–3 minutes (section 17).
3. 330 of the 356 trees in the park have an assumed species; their leaf area, leaf angle and leaf habit are defaults.
4. Tree data is from 2017–2019; seven to nine growing seasons have passed.
5. Wind shelter ignores flow deflection, channelling and gusts.
6. Fixtures, mowing and sprinkler schedules are not published; those outputs are probabilities.
7. Busyness, dogs, the rotunda and party shares are assumptions until check-ins refit them.
8. Below half a degree of sun elevation the far horizon is not reliable, so last sun is not resolved past that point.
9. The 3D building model's licence for public redistribution is not yet confirmed.
10. Check-ins and survey answers are self-selected, not a representative sample; until there are enough, forecast accuracy is unknown.

22. Accessibility statement

Target: WCAG 2.2 level AA. Every control is reachable and operable by keyboard with a visible focus indicator; dialogs trap focus, close with Escape and return focus; camera movement has button and keyboard alternatives to dragging; pins can be placed by typing a grid reference or using device location; charts carry text descriptions and multi-series lines differ by dash pattern as well as colour; every coloured category has a text label; text contrast is at least 4.5:1; targets are at least 24 × 24 px (44 px on touch screens); zoom is never blocked; status messages are announced to screen readers; reduced-motion and increased-contrast settings are respected; optional high contrast, 125–175% text and controls, and the Atkinson Hyperlegible typeface.

Automated checks (axe-core 4.10, WCAG 2.0, 2.1 and 2.2 A/AA rules) report no violations on the main view, a spot profile, every dialog and high-contrast mode. Automated tools catch a minority of issues; testing with screen readers and disabled users is still to come. The 3D view cannot be fully described to a screen reader; its information is available as text through the text summary, the spot profile and the accessible spot finder. Report barriers with Suggest a feature ("Accessibility barrier").

23. References

1. Allen RG, Pereira LS, Raes D, Smith M (1998). Crop evapotranspiration. FAO Irrigation and Drainage Paper 56. fao.org
2. Appleton J (1975). The Experience of Landscape. Wiley.
3. Bradley RA, Terry ME (1952). Rank analysis of incomplete block designs I. Biometrika 39: 324–345.
4. Brier GW (1950). Verification of forecasts expressed in terms of probability. Monthly Weather Review 78: 1–3.
5. Bröde P et al. (2012). Deriving the operational procedure for the Universal Thermal Climate Index (UTCI). International Journal of Biometeorology 56: 481–494.
6. Campbell GS (1986). Extinction coefficients for radiation in plant canopies calculated using an ellipsoidal inclination angle distribution. Agricultural and Forest Meteorology 36: 317–321.
7. Campbell GS (1990). Derivation of an angle density function for canopies with ellipsoidal leaf angle distributions. Agricultural and Forest Meteorology 49: 173–176.
8. Chianucci F et al. (2018). A dataset of leaf inclination angles for temperate and boreal broadleaf woody species. Annals of Forest Science 75: 50.
9. Churchill SW, Usagi R (1972). A general expression for the correlation of rates of transfer and other phenomena. AIChE Journal 18: 1121–1128.
10. de Freitas CR, Scott D, McBoyle G (2008). A second generation climate index for tourism (CIT). International Journal of Biometeorology 52: 399–407.
11. Department of Transport, Welsh Office (1988). Calculation of Road Traffic Noise. HMSO.
12. Fanger PO (1970). Thermal Comfort. Danish Technical Press.
13. Fretwell SD, Lucas HL (1970). On territorial behavior and other factors influencing habitat distribution in birds. Acta Biotheoretica 19: 16–36.
14. Gehl J (1987). Life Between Buildings. Van Nostrand Reinhold.
15. Hall ET (1966). The Hidden Dimension. Doubleday.
16. Heisler GM, DeWalle DR (1988). Effects of windbreak structure on wind flow. Agriculture, Ecosystems and Environment 22/23: 41–69.
17. Hersbach H et al. (2020). The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society 146: 1999–2049.
18. Hilpert R (1933). Wärmeabgabe von geheizten Drähten und Rohren im Luftstrom. Forschung auf dem Gebiet des Ingenieurwesens 4: 215–224.
19. Hoerl AE, Kennard RW (1970). Ridge regression. Technometrics 12: 55–67.
20. ISO 7726:1998. Ergonomics of the thermal environment: instruments for measuring physical quantities.
21. Kasten F, Young AT (1989). Revised optical air mass tables and approximation formula. Applied Optics 28: 4735–4738.
22. Lindberg F, Grimmond CSB (2011). The influence of vegetation and building morphology on shadow patterns and mean radiant temperatures in urban areas. Theoretical and Applied Climatology 105: 311–323.
23. McKelvey RD, Palfrey TR (1995). Quantal response equilibria for normal form games. Games and Economic Behavior 10: 6–38.
24. Meeus J (1998). Astronomical Algorithms, 2nd ed. Willmann-Bell.
25. Meinel AB, Meinel MP (1976). Applied Solar Energy. Addison-Wesley.
26. Pisek J, Adamson K (2020). Dataset of leaf inclination angles for 71 different Eucalyptus species. Data in Brief 33: 106391.
27. Steadman RG (1994). Norms of apparent temperature in Australia. Australian Meteorological Magazine 43: 1–16.
28. Whyte WH (1980). The Social Life of Small Urban Spaces. Conservation Foundation.
29. World Health Organization (2002). Global Solar UV Index: A Practical Guide.
30. World Meteorological Organization (2017). WMO Guidelines on the Calculation of Climate Normals (WMO-No. 1203).
31. ECMWF. IFS documentation. ecmwf.int
32. Australian Bureau of Statistics. 2021 Census G04; Births, year and month of occurrence. abs.gov.au
33. State of Victoria. Vicmap Vegetation – Tree Urban; Vicmap Elevation; Vegetation Cover for metropolitan Melbourne 2018; Annual Average Daily Traffic Volume 2019. CC BY 4.0.
34. City of Yarra. Domestic Animal Management Plan 2021–2025; significant trees; street and park trees; permit parking zones (CC BY 4.0); Edi Gardens; "Water upgrades at Edinburgh Gardens to save millions of litres" (2021); General Local Law.
35. Open-Meteo. Weather, ensemble, previous-runs, historical and air-quality APIs. open-meteo.com, CC BY 4.0.
36. Copernicus Atmosphere Monitoring Service (CAMS) global forecasts.
37. Australian Government. National Public Toilet Map. toiletmap.gov.au, CC BY 4.0.
38. Standards Australia. AS 1428.1 Design for access and mobility.
39. Victorian Department of Health. Heat health thresholds; epidemic thunderstorm asthma forecast. health.vic.gov.au
40. EPA Victoria. AirWatch air quality categories. epa.vic.gov.au
41. Country Fire Authority. Fire Danger Ratings and Total Fire Bans, Central district. cfa.vic.gov.au
42. Melbourne Pollen. Grass pollen categories. melbournepollen.com.au
43. Nager.Date. Public holiday API. date.nager.at
44. Heritage Council of Victoria, Victorian Heritage Database. Fitzroy Memorial Rotunda (103220); Edi Gardens.
45. City of Melbourne. Pedestrian counting system. data.melbourne.vic.gov.au
46. W3C (2023). Web Content Accessibility Guidelines 2.2. w3.org/TR/WCAG22
47. OpenStreetMap contributors. openstreetmap.org/copyright, ODbL.

Product IDB3D001 · Methods revised 15 October 2026. Back to the forecast.