What the error bars mean

Where the bars come from

Every time you press Auto-detect shading, we look back over the past few months of inverter readings. For each day, we compare what your panels actually produced against what the calculator would have predicted for that day’s real weather. Most days the prediction is close. Some days it overshoots, others it undershoots.

We sort all those days into five groups by how cloudy they were (almost clear, mostly clear, mixed, mostly cloudy, overcast). Inside each group we look at:

• How far off the predictions were on average — this becomes a calibration adjustment for next time.
• How scattered the misses were — this becomes the width of the error bars for next time.

How the calculator uses them

When you click Calculate, for every day in the window the calculator looks at how cloudy the day is (forecast or actual, depending on the date), then does two things:

1. Nudges the prediction toward what your panels usually do in this kind of weather. We tend to slightly over-predict on sunny days and under-predict on heavily overcast ones, so the nudge corrects for that.
2. Sizes the error bar based on how unreliable predictions have been for this cloud level. Sunny days are easy and get tight bars. Patchy-cloud days are hardest and get the widest bars.

Why even past days have error bars

You might expect a perfectly known past day to have no uncertainty — we know the weather that actually happened. But predicting solar production from cloud cover is still an imperfect art: the same “70 % cloudy” can mean a steady grey blanket or fast- moving cumulus, and the solar output differs. The error bars on past days reflect that intrinsic difficulty.

So a past-day bar means “here’s how much we typically miss by when the weather was this cloudy” — not weather-forecast uncertainty (there isn’t any once it’s happened), just model uncertainty.

Reading the per-day table

• Total — the most-likely kWh for the day, after the calibration nudge.
• Low — a realistic worst case. About 1 day in 6 like this one comes in at or below this number.
• High — a realistic best case. About 1 day in 6 like this one comes in at or above this number.
• Forecast uncertainty — the give-or- take, in kWh. Roughly two days in three should land inside Total ± this much.
• Measured — the actual production from your inverter, when there is one. This column only appears if you have an inverter connected.

The headline number above the chart

The total at the top of the chart (e.g. “2255 ± 59 kWh”) is the sum of all the daily predictions, with an honest ± figure that accounts for the daily error bars. It’s smaller than just adding up every day’s ± because daily forecast errors tend to cancel each other out across a long window — some days the weather lands cloudier than expected, others sunnier, and over weeks the differences roughly even out.

Why this matters

A single number can mislead. If you’re sizing an off-grid battery, you care about the Low column — the realistic worst case. If you’re forecasting export earnings, you care about the Total. And if you want to know whether a tariff would have covered your bill, you want confidence that your real result will fall inside the band.

Without the band, every number is a guess presented as a fact. With it, you can plan around the range of likely outcomes — which is what actually happens in practice.

New layouts: the band is a starter estimate

Brand-new layouts (no inverter connected, or just haven’t run auto-detect yet) use a reasonable default band based on weather-forecast skill. It works, but it doesn’t know anything specific about your panels. Connecting an inverter and running auto-detect upgrades to the calibrated band — usually much tighter on clear days and a bit wider on cloudy days.

The more measured days the system has to learn from, the better the bars get. Running auto-detect again every few months keeps the calibration fresh.

A note for the technically curious

When you have measured data for past days, the chart header shows a small “χ²/ν” number. This is a statistical check on whether the error bars are the right size:

• Around 1.0 — bars are honest. Measured days land inside the band about as often as expected.
• Much larger than 1.0 — bars are too narrow; the model is over-confident.
• Much smaller than 1.0 — bars are too wide; the model is over-cautious.

The number gets more reliable as you accumulate more measured days. Running auto-detect again after a few more months will refit the bars on a bigger dataset and should keep this near 1.0.

See also the Forecast confidence help page for the simpler default band used when no inverter history is available yet.