NEM Dashboard · week-ahead forecast

Price Lab

Australia's week-ahead wholesale price forecast

Every Sunday this page publishes a seven-day forecast of the wholesale spot price for each NEM region. The following Sunday each forecast is scored against what actually cleared. Predictions are never revised, so the record accumulates in public, including the weeks the models lose.

Project thesis · what the Price Lab is testing Read the thesis

Most public commentary on wholesale electricity prices is explained after the fact, once the number is already known. The Price Lab tests the harder, honest version of that claim: publish a call every Sunday, before the week happens, and score it against what actually cleared the following Sunday. A forecast that is only ever discussed in hindsight is not really a forecast.

Six models compete on the same public price and temperature history, ordered from a naive baseline through to a neural network, and each has to beat the ones below it to justify its own complexity. This is a teaching model built on public NEM data, not a trading signal: it is not revised after publication, and it is not financial advice.

The point is not to claim a superior forecast. It is to make the comparison itself transparent: the same baseline, the same scoring, the same archive of every past call, so a reader can see exactly how much a more elaborate model actually buys over the simplest possible guess, week after week.

Weekly data refresh Runs Mondays at 09:00 Australia/Sydney · settled to –

Price Character NEM

Average price

–

$/MWh · last 7 days

Lowest interval

–

$/MWh · 30-minute

Highest interval

–

$/MWh · 30-minute

Below zero

–

share of intervals paid to stop

The Week Ahead

–

Forecasts run from Sunday to Sunday. Each one is fixed at its origin and never revised. Solid black is what actually cleared.

Show models Same as last week baseline – Median of last 4 weeks – Last week + temperature – Random forest – Gradient boosting – Neural network –

The Six Models

Region NEM

Ordered cheapest to most elaborate. Each rung has to beat the ones below it to justify existing, and the scores shown are for the region above.

Performance

– Region NEM

Two horizons. Last week is the most recent forecast whose week has fully settled. Average to date covers every completed week since the record began, and is the number that carries weight, because one week can be won by luck.

What "Skill" Means

Skill compares a model against the baseline of simply assuming this week repeats last week. +19.7% skill means the model's average error was 19.7% smaller than the baseline's, not that it was 19.7% away from the real price. A negative number means the model did worse than doing nothing clever at all. The baseline scores 0% against itself by definition.

Model	Last week			Average to date			Worst single interval
Model	Mean error	Median error	Skill	Mean error	Median error	Skill	Worst single interval

Mean and median are both shown on purpose. A single spike interval can dominate the mean, so the median says how the forecast behaves on an ordinary half hour while the mean says what the spikes cost. Reporting only one would be choosing which story to tell.

Eight-Week Performance Trend

The actual settled price against each model's average predicted price, week by week, over the last eight completed weeks. The error and skill numbers above are derived from this comparison; this is that same comparison made visible over a longer stretch than one week.

What the Model Actually Learned

Region NEM

The entire fitted part of this system is two dollars-per-degree numbers per time-of-day band. Here they are, in full.

Heating and cooling are fitted separately, against an 18 °C base, because demand responds to temperature in a V rather than a line: in summer a hotter week is a dearer week, but in winter a colder week is a dearer week. The first version of this model used a single coefficient and measured the morning response as +$11.57/°C in summer and −$12.74/°C in winter, which averaged out to +$1.47 and made the model almost useless. Splitting the driver lets each side keep its own sign.

Band	Hours	Cooling $/MWh per °C above 18	Heating $/MWh per °C below 18	Pairs used	Why it matters

Where the Price Has Been

Last 90 days Region NEM

Daily average price, with the shaded band showing each day's cheapest and dearest half hour. The gap between them is the story: it is routinely wider than the average itself.