TLZ EnergyNSW1 price chart & forecast
Read it like a trading chart. Intraday shows the 48 half-hours of one day; 1D / 1W / 1M / 1Y show the settled price either as candles (open, high, low, close of the half-hourly prices in each period) or as a line of the period's average price. The blue line and band are the TLZ forecast average and its 80% interval average; the dashed line is AEMO's 12:30 pre-dispatch. The lower pane shows how often settled prices landed inside our interval.
How the live forecasts have performed
Every day since the live pipeline started, compared with what actually settled. Each forecast is published before the day it covers and is never revised afterwards; the model is retrained weekly, only on days that have already settled. Click a row to load that day in the chart above.
Daily mean absolute error ($/MWh)
Daily results
Twelve months of rolling, out-of-sample evaluation
Each day was forecast at 12:35 the day before using only information available at that time, then scored against the settled price. We compare against AEMO's pre-dispatch price and against a full replication of a published academic method (Q-QRA, Cornell, Dinh & Pourmousavi, 2023) run on the same data.
Interval score by method (lower is better)
Winkler interval score for the 80% interval.
Show table
Coverage by hour of day
An 80% interval should contain the settled price 80% of the time — in every hour, not just on average. Conformal calibration is what keeps every hour close to the target.
Monthly interval score (lower is better)
Built for calibration, not just accuracy
Most price forecasts are a single line. Trading and storage decisions depend on the range of outcomes, so we forecast quantiles and then calibrate them so that the stated probabilities hold up.
Market data
AEMO NEMWEB public data, ingested automatically.
- 12:30 pre-dispatch prices for NSW, QLD, VIC and SA
- Forecast demand, available generation, reserve margin, wind and solar
- Recent settled prices and demand
- Recent pre-dispatch errors
Quantile model
Gradient-boosted quantile regression on market and pre-dispatch price-sensitivity features.
- 10th, 50th and 90th percentiles for all 48 half-hours
- Variance-stabilised price scale handles negative prices and spikes
- Retrained weekly as new settled data arrives
Conformal calibration
Adaptive, hour-by-hour calibration layer.
- Separate lower and upper corrections for each hour of day
- Recent errors weighted more heavily as the market evolves
- Keeps interval coverage at its stated level
Coverage
Get the forecasts in your workflow
Daily CSV and JSON feeds, full quantile curves and historical backtests are available on request.