This note measures how price risk is shared between the five regions of the National Electricity Market, and how that sharing follows the interconnector network.
- Prices move in a fixed order. 91% of the lead–lag signal between regions is one consistent ordering: Queensland first, then New South Wales, Victoria and Tasmania, and South Australia last. That is the order of longitude, consistent with the daily solar cycle reaching each region at its local time, and the ordering doubled in strength from FY22 to FY26.
- Pairwise couplings are enough. A model with only pairwise couplings accounts for 99.3% of the shared information in five-region co-spiking, and its four strongest couplings are the four interconnectors.
- A binding import limit separates extremes. South Australia and Victoria are directly connected, yet 67% of South Australia's top-1% half-hours occur without Victoria's, every one of them after an SA import limit has bound. Elsewhere, joint extremes fall with each link between two regions: 0.48 for neighbours, 0.19 three links apart.
- Disturbances spread smoothly over the network. Price disturbances put 46% of their energy at the coarsest network scale, against 33% for unstructured noise. Generator trips are the most network-wide; interconnector swings the most local.
The data are AEMO 5-minute regional prices for New South Wales, Queensland, South Australia, Tasmania and Victoria from 1 October 2021 to 31 May 2026: 490,751 intervals per region, with the June 2022 market suspension excluded. The interconnectors form a tree. Queensland, New South Wales and Victoria are linked in a line, and South Australia and Tasmania are each linked only to Victoria. Any two regions are one, two or three links apart.
share of the lead–lag signal that is a consistent order (gradient). The other 9% is circulation (curl).
share of five-region co-spiking structure accounted for by pairwise couplings alone.
rank correlation between joint extreme prices and the number of links apart, over the ten region pairs.
Joint extremes fall with every link between two regions, and South Australia's mostly come after an import limit binds
For each region we took the highest 5-minute price in every half hour and marked the top 1% of half-hours as extreme. The thresholds are $605/MWh in New South Wales, $635 in Queensland, $694 in South Australia, $450 in Tasmania and $476 in Victoria, which gives each region 810–819 extreme half-hours.
For two regions, the co-exceedance χ is the number of half-hours in which both are extreme, divided by the geometric mean of their separate counts. It is 1 if their extremes always coincide and near 0 if they never do. Shifting one region's record by a random whole number of days gives 0.02.
The ordering by distance is complete. Every directly connected pair has a higher χ than every pair two links apart, and every pair two links apart has a higher χ than both pairs three links apart. The group means are 0.48, 0.27 and 0.19. The rank correlation between χ and links apart is −0.93, the largest magnitude possible with tied distances (95% day-block bootstrap interval −0.93 to −0.75). The ordering is equally complete at the top 5%, 2% and 0.5% of half-hours, and the correlation is between −0.64 and −0.88 within each financial year. The margins at the group boundaries are small: SA–VIC (0.33) against QLD–VIC (0.32), and SA–TAS (0.20) against QLD–SA (0.20).
Note 01 found the same gradient for spike onsets: an onset raises the 30-minute onset probability 8.9-fold in a directly connected region, 3.9-fold two links away and 2.6-fold three links away.
A straight line, χ ≈ 0.62 − 0.16 × links, leaves structured residuals. Pairs that include Queensland sit above it: NSW–QLD by 0.14, QLD–SA and QLD–TAS by 0.04. Pairs that include South Australia sit below it, except QLD–SA: SA–VIC by 0.13, SA–TAS by 0.11 and NSW–SA by 0.05. Each of these six has a 95% bootstrap interval that excludes zero.
SA–VIC is directly connected but has the lowest χ of the four wired pairs. Of South Australia's 810 extreme half-hours, 540 (67%) occurred without Victoria in its own top 1%. In all 540, an SA import interconnector had been at its limit during the preceding hour, and South Australia's half-hour maximum was a median 23 times Victoria's. In the 270 half-hours when both were extreme, an SA import limit had bound in 26%, and the two prices were nearly equal (median ratio 1.01). Queensland and Tasmania show the same pattern at lower shares: 39% of Queensland's extreme half-hours occurred without New South Wales, and 50% of Tasmania's without Victoria, with the import limit binding in 100% and 97% of them.
An interconnector couples two prices only while it is below its limit. Once it binds, the two regional prices are set separately and the importing region's extremes become local. For South Australia, that is the usual case at the top 1%.
Pairwise couplings capture 99.3% of the shared information in five-region co-spiking; the four strongest are the four interconnectors
Pair statistics can mislead when three or more regions spike together, because two regions can be correlated through a third. To separate direct from indirect association, we fitted a pairwise maximum-entropy model (an Ising model) to 80,975 half-hours on 1,687 days. A region is spiking in a half-hour if its price exceeded $300/MWh in any 5-minute interval of it.
The model gives each region a field h and each pair a coupling J. The probability of a spike pattern s (+1 spiking, −1 not, for each region) is proportional to exp(Σ hisi + Σ Jijsisj). Fitted exactly over all 32 possible patterns, it reproduces each region's spike rate (3.8–7.1% of half-hours) and each pair's correlation, and adds no other structure.
The four largest couplings are the four interconnectors: NSW–QLD 1.00, SA–VIC 0.98, TAS–VIC 0.87 and NSW–VIC 0.73, each with a 95% day-block interval above 0.62. The largest coupling without a direct link is QLD–SA, 0.37 (0.30–0.43), the same pair that sits above the distance line in Chapter 01. QLD–VIC shows the difference between correlation and coupling: their spike correlation is 0.46, but their coupling is 0.03 (−0.06 to 0.13). The model attributes their co-spiking to the path through New South Wales.
SA–VIC is the second-strongest coupling at the $300 level used here, while its top-1% co-exceedance is the weakest of the wired pairs. The two results are consistent with Chapter 01: the SA import limit separates the two prices at South Australia's highest prices, not at $300.
The pairwise model is sufficient to within 1%. The five spike indicators share 0.467 bits of information per half-hour beyond what their separate rates imply (the multi-information). The pairwise model accounts for 99.3% of it (95% interval 99.0–99.5%). The remaining 0.68% (0.50–1.05%) is three-way and higher interaction. It is real, since sampling alone would leave about 0.03%, but it is small.
The all-region pattern shows the size of the pairwise effect. All five regions spiked together in 1,172 half-hours (1.4%). Independent regions with the same spike rates would produce that pattern 0.03 times in the whole period; the pairwise model predicts 1,274. For simulating joint spikes, five rates and ten couplings are enough.
Price disturbances are smooth over the network: 46% of their energy is at the coarsest scale, against 33% for noise
We selected 21,961 five-minute intervals in which a physical disturbance coincided with a price move of at least $50/MWh from the trailing-hour mean in some region. The disturbances are generator trips (at least 200 MW lost within 15 minutes; 14,216 intervals), large interconnector swings (the top 1% of 15-minute changes in a region's net interchange, at least 575 MW; 3,027) and large ramps (the top 1% of 15-minute changes in demand, at least 426 MW, or in price; 7,124). An interval can belong to more than one group. Each disturbance is the vector of the five regions' price deviations from their trailing-hour means, on the asinh price scale.
The interconnector graph, with each link weighted by its nominal transfer capacity, defines five patterns of variation across the regions: one in which all five move together, and four contrasts of increasing roughness, from a smooth gradient along the chain to one region moving against its neighbour. These are the eigenvectors of the graph Laplacian. Spectral graph wavelets group them into four scales, from global to local, and divide each disturbance's energy among the scales.
The median disturbance puts 46% of its energy in the global scale, against 33% (95% band 33.0–33.5%) for unstructured noise, and less in the local scale (14% against 21% of total energy). After removing the all-regions-together component, the remainder is still smooth over the wires: its energy-weighted graph frequency is 0.27 of the maximum, against 0.37 (0.369–0.379) when the five values are shuffled among the regions. Neighbouring regions deviate together more than non-neighbouring ones.
Generator trips are the most global (median 0.49; 95% interval 0.485–0.494). Large ramps follow (0.42), and interconnector swings are the most local (0.41; 0.402–0.419). The global share was 0.44–0.47 in every financial year.
Interpretation: a trip removes supply from the synchronised mainland system, and the response (frequency control, reserve, rebidding) draws on generators in several regions, so its price effect is shared. A large interconnector swing changes the flow between two regions and, near a limit, separates their prices, so its effect is more local. The measurement ranks the three triggers; it does not isolate these mechanisms.
Prices move in a fixed order, Queensland first and South Australia last; 91% of the lead–lag signal is that order
For two price series, the Lévy area is the signed area enclosed by their joint path over a window. It is positive when the first series tends to move before the second, needs no model of either series, and covers every timescale shorter than the window. We computed it for every pair of regions in each of 240 weeks, on asinh-transformed 5-minute prices with each region's weekly increments scaled to unit total variation, and averaged over weeks.
The ten average areas form a flow between the five regions. A flow on a network in which every pair is connected splits into two orthogonal parts (the combinatorial Hodge decomposition). The gradient part gives each region a score and sets each pair's area to the difference of their scores: a consistent ordering, or potential. The curl part circulates around triangles of regions (A before B, B before C, C before A) and has no first or last. With every pair connected there is no third part, so the two shares add to 100%.
The gradient holds 90.7% of the energy (95% week-bootstrap interval 88.7–92.3%) and the curl 9.3%. A random flow on five regions would put 40% in the gradient, because the gradient part spans 4 of the 10 pair dimensions. Scrambling the order of each region's increments within each week gives 41% on average (95th percentile 75%). The scores are Queensland +0.031, New South Wales +0.008, Victoria −0.009, Tasmania −0.009 and South Australia −0.020. Queensland is first and South Australia last in all 4,000 week-bootstrap resamples. The largest single area is Queensland over South Australia (0.062).
The direction is stable from week to week. The dominant lead–lag pattern has the same sign in 225 of 240 weeks (94%; binomial p = 3 × 10−49), and Queensland has the highest score in 77% of individual weeks. The weekly areas are small: the typical weekly rotation is 0.13 times that of the same increments with their timing scrambled, because the five prices mostly move together. The ordering is a consistent bias on top of strong co-movement.
The order follows longitude. From east to west the capitals are Brisbane (153.0°E), Sydney (151.2°), Hobart (147.3°), Melbourne (145.0°) and Adelaide (138.6°). The lead scores fall in that order, with Victoria and Tasmania tied. A cross-spectral phase measurement on the same prices gives the same direction at the two ends of the chain. At the 24-hour harmonic, South Australia's price cycle trails New South Wales's by 63 minutes over 2021–26 (coherence 0.86) and by 48 minutes in the 12 months to May 2026 (coherence 0.91). Solar time in Adelaide runs 50 minutes behind Sydney. Queensland leads New South Wales by 1 and 10 minutes in the two windows at 24 hours and by 17–18 minutes at the 12-hour harmonic, against a solar offset of 7 minutes. Victoria and Tasmania are not consistently ordered by phase either: Tasmania trails New South Wales by 28 minutes over 2021–26 but leads by 16 minutes in 2025–26.
Clock time does not produce the order. From October to April New South Wales observes daylight saving and Queensland does not, so demand that follows the clock arrives an hour earlier in New South Wales, in market time. Queensland still leads New South Wales in those weeks (mean area +0.007, positive in 65% of 126 weeks), at about a third of its lead in standard-time weeks (+0.021, positive in 81% of 104 weeks).
Interpretation: the order is the daily solar cycle, meaning the midday price trough and the evening ramp after sunset, reaching each region at its local solar time. Removing each region's average daily price profile leaves 30% of the lead–lag magnitude, still with Queensland first and 90% in the gradient. That remainder comes from price moves away from the average day.
The ordering doubled in strength between FY22 and FY26. The projection of each year's average area matrix on the dominant lead–lag pattern rose from 0.081 (90% interval 0.063–0.100) in FY22 to 0.164 (0.135–0.195) in FY26, and the gradient share was 84–94% in every year. The sign was positive in 89–98% of weeks in each year. Queensland's score rose from +0.020 to +0.039 and South Australia's fell from −0.013 to −0.028. From FY24 the yearly strength lies outside the range produced by scrambled timing; in FY22 and FY23 it lies within it, and the evidence for those two years rests on the consistency of the sign.
Interpretation: deeper midday price troughs from rooftop and utility-scale solar increase the amplitude of the daily price cycle, and a larger cycle arriving at different times produces a larger area. We have not tested this against measured solar output here.
What this means for risk, forecasting and modelling
Extreme prices coincide in about half of the cases for directly connected regions (χ 0.48) and a fifth for regions three links apart (0.19). The exception is SA–VIC: two-thirds of South Australia's extremes occur behind a binding import limit.
Five spike rates and ten couplings account for 99.3% of five-region co-spiking structure. Simulating joint spikes needs the couplings; three-way and higher terms add 0.7%.
Price moves reach Queensland first and South Australia last, about an hour apart at the daily cycle, and the ordering has doubled in strength since FY22. Earlier regions' current prices are informative inputs for later regions' forecasts.
WEFT forecasts all five regions together from each region's recent 5-minute history (price, net interchange and more) and its interconnector headroom and binding time, so the couplings, the binding effect and the arrival order are visible to the model.