Live price intelligence, NEM
18 months. Five regions. Every five minutes. Tested against prices the model had never seen.
For every 5-minute interval across the NEM, the tool publishes the most likely price, the spike probability and an 80% confidence range — and then holds itself to account. We ran it continuously through two summers, price-cap events and ordinary trading days, and scored every forecast it made against the settled price. Here is the record.
Wholesale prices forecast every five minutes, day ahead
A full picture for every interval, not a single number. The nowcast anchors the horizon; the same control logic looks ahead whether the saving comes from spot, TOU or demand charges.
Most likely price
The expected price for each interval, from conditions the model has learned to associate with it.
Spike probability
How likely a spike is: the event that matters most to a plant with flexible load. When the tool warns, it is right — 93–99% of alerts were verified real spikes.
Realistic range
The 80% interval the price is expected to fall within. Wide enough to be honest about uncertainty. The nowcast is built from live market data and anchors the rest of the horizon.
What 18 months of public testing showed
Every forecast in the window was scored against the settled price at the same 5-minute interval — no cherry-picking, no regime filtering, no looking ahead.
The market it is up against routinely swings from $0 to hundreds — and in this window threw genuine price-cap events at the model (up to ~$20,000/MWh). Region-by-region results below.
Results by region, 18 months (wMAPE): click to expand
| Region | Nowcast: our tool | Nowcast: AEMO look-ahead | Day ahead: our tool | Day ahead: AEMO (2025) |
|---|---|---|---|---|
| QLD | 16.7% | 35.1% | 61.0% | 160.2% |
| NSW | 20.4% | 54.5% | 69.0% | 188.3% |
| VIC | 22.4% | 82.1% | 79.5% | 87.6% |
| SA | 24.7% | 58.1% | 89.9% | 119.4% |
| TAS | 14.8% | 47.3% | 43.4% | 65.2% |
| Average | 19.8% | 55.4% | 68.6% | 124.1% |
wMAPE = average miss as % of actual price. 18-month window: 1 Jan 2025 – 30 Jun 2026, all 5-min intervals, tool scored 5-min vs 5-min. AEMO look-ahead scored from published pre-dispatch (~28 min look-ahead at 30-min resolution). AEMO day-ahead figures are the 2025 published pre-dispatch 12–36h series from the project’s earlier backtest.
Why the long run matters
Past results with one quiet year are a snapshot. This is a track record: two summers, price caps, interconnector outages and calm periods all in the same 18-month window, and the accuracy held.
- Proven under the hard cases. The window included genuine scarcity events up to ~$20,000/MWh. In normal trading the median miss is ~$5; when the market did spike, the tool’s alerts were right 93–99% of the time — no crying wolf.
- Honest by construction. The published 80% range contained the actual price 84–91% of the time — on target, not over-padded.
- Beats the market’s own numbers. AEMO’s look-ahead, published by the market operator for the same task, shows roughly two to five times the error at the nowcast.
- Runs daily, retrains in the background: downloads fresh data and publishes the next day’s forecast with confidence ranges, automatically.
Why it matters, even if you stay on a tariff
Forecasting turns a simulation into control. Backtesting against real settlement data shows what would have been saved; forecasting makes that logic useful forward: deciding to pump now or wait, to hold biogas or generate, to charge a battery or discharge. A tariff-optimised plant still needs to look ahead — TOU windows, demand-peak risk, solar and inflow all have a trajectory.
Sources: AEMO NEM dashboard, AEMO pre-dispatch, BOM weather.
The first step is a plant assessment
The forecast is the price data behind the simulation. For your plant, the assessment, modelled against real market and tariff data, establishes whether there is a case and which of the three routes it falls under.