This note takes the NEM's daily cycle of price, reserve, demand and wind as a reference path and measures how prices leave it: in what order price and reserve move before a spike, how high prices are grouped in time, and how 5-minute price changes are distributed.
- Before a spike, reserve falls first. Over the two hours before 922 storms (episodes peaking above $1,000), median reserve fell from 0.70 to 0.40 of its calm level while the median price stayed below $300 until the last 10–15 minutes. Under the fitted 5-minute dynamics, a route to the cap that raises price first costs 5.3 times as much as the least-action route, on which reserve falls first. On ordinary days the order is the reverse: mainland price turns 5–60 minutes before reserve.
- High prices come in fewer, longer runs than their autocorrelation implies. Prices above $300 form 32% fewer runs than the same prices rearranged with the same spectrum, and each run lasts longer: 33 minutes against 22.
- Jumps carry most of the 5-minute variance. At most price levels a minority of large moves carries 83–99% of it, no price level at $300 or above is stable, and 14% of prices at or above $5,000 sit exactly on the cap.
- Price rings after a disturbance. A damped oscillation with a period of about 15 minutes appears in every region and year: about a third of a 5-minute move reverses in the next interval.
The daily cycle itself is familiar: prices rise into the evening peak and fall overnight, and each day's states trace one loop. Here it serves as the reference against which these departures are measured.
The data are AEMO's 5-minute dispatch data for all five regions, from 1 October 2021 to 31 May 2026: regional price, demand, available generation, reserve (available generation minus demand) and wind output. Intervals in the June 2022 market suspension and intervals under administered pricing are excluded. Price is analysed on an asinh scale, asinh(price ÷ 30), which is close to linear between −$30 and $30 and logarithmic beyond a few hundred dollars, so that the −$1,000 floor and the $15,100–$20,300 cap fit on one axis. Several results come from our earlier analyses of the same data; the text marks them and the side notes give their methods.
the action of a price-first route to the price cap, relative to the least-action route, on which reserve falls first.
runs above $300 per region-day, observed against the same prices rearranged with the same autocorrelation (1.9 against 2.8).
period of the damped oscillation that follows a price disturbance, in all 25 region-years (damping ratio 0.47).
On ordinary days, mainland price turns 5–60 minutes before reserve
Each day's 288 five-minute states in price, reserve, demand and wind trace one loop: the daily cycle. An earlier analysis measured it with persistent homology, with time of day not an input, for every day in New South Wales and South Australia. Virtually every day has one persistent loop, median size 0.35 standard deviations: 0.26 at the June solstice and 0.45–0.49 in high summer. There is no second, independent loop, such as an annual cycle would add. Loop size is unrelated to the day's price range (correlation +0.03 to +0.04): the largest loop in the record, New South Wales on 14 December 2025 (2.05), was a Sunday on which the price never exceeded $134. A spike appears as an excursion away from a loop of ordinary size.
The loop has a direction. Figure 1 plots, for each 5-minute slot of the day, the median reserve and the median price over all valid days. Reserve is divided by the region's median reserve in calm intervals (prices at or below $100), so that the five regions share one axis. In all five regions the median day runs counter-clockwise: price rises before reserve falls, and falls before reserve recovers. Cross-correlating the two median profiles, price leads reserve by 25 minutes in New South Wales and Victoria (correlation 0.98 and 0.89), 60 minutes in South Australia (0.99) and 5 minutes in Queensland (0.96). Tasmania's reserve barely co-moves with its price (0.24). On the median day the price peaks at 18:00–18:30 in every region, and mainland reserve reaches its minimum 0–40 minutes later.
Interpretation. Price responds to expected scarcity before physical reserve reaches its minimum. That ordering is consistent with offers being repriced ahead of the evening peak on the basis of forecasts, rather than prices following reserve mechanically; the data show the ordering, not its cause. Because loop size does not predict spike days, the loop describes the ordinary daily cycle. Spikes are departures from it, and the rest of this note measures those departures.
Prices above $300 arrive in 32% fewer and longer runs than the same prices rearranged
For one day's price series, the number of separate runs above a threshold is the Euler characteristic of the set of times above that threshold; sweeping the threshold gives the day's Euler-characteristic curve. Averaged over 8,430 region-days, there are 1.90 separate runs above $300 per region-day.
Whether that count reflects how prices are arranged in time, rather than which prices occur, can be tested with surrogate days built from exactly the same 288 prices, reordered so that the power spectrum, and therefore the autocorrelation including the daily cycle, matches the original (the iterative amplitude-adjusted Fourier transform). The surrogates have 2.79 runs above $300 per region-day. The observed days have 32% fewer runs, and each run is longer: 6.5 intervals (33 minutes) against 4.5 (22 minutes). Observed days have fewer runs than their surrogates at every one of the 48 thresholds from $20 to $16,000.
The same holds across days. Figure 3 shows South Australia's prices as a carpet: one column per day, one row per 5-minute interval of the day. A surrogate carpet built from exactly the same prices, keeping the average daily profile and each day's level and rearranging the deviations from them with their own two-dimensional spectrum, has 3,687 separate islands above $300, against 2,347 observed: 36% fewer observed, with a mean island size of 9.7 intervals against 6.2. The observed carpet also has fewer islands above $1,000 (407 against 660) and above $5,000 (217 against 351).
Interpretation. Once the price is above $300, it stays there longer than its autocorrelation alone implies. The persistence is nonlinear: a linear model of price fitted to the whole series, which reproduces the spectrum, still breaks high prices into too many short runs. For forecasting, spike duration is a property of the episode, and the chance that the next interval stays above $300 is higher than a linear model gives. For risk, the same number of high-price intervals arrives in fewer, larger blocks. Separately, the mean number of runs above $300 per region-day fell from 3.36 in FY22 to 0.58 in FY26.
Before a spike the order flips: on the least-action route to the cap reserve falls first, and a price-first route costs 5.3 times as much
The state of a region can be reduced to two numbers: reserve relative to its calm median (ρ) and price. An earlier analysis estimated, from every valid 5-minute increment pooled across the five regions, the average move (drift) and the covariance of moves in each cell of this plane. It then found the path from the calm state to the price cap that requires the least improbable sequence of moves: the minimum-action path of large-deviation theory.
Along that path, reserve falls first. Starting from ρ = 0.88 at $57, reserve falls to ρ = 0.35 while the price stays below $350; the price then climbs to $1,500 at ρ = 0.27 and reaches the cap at ρ = 0.04. Its action is 3.80. A straight diagonal between the same end points costs 8.65 (2.3 times as much). A route that lowers reserve all the way before raising price costs 12.76 (3.4 times), and a route that raises price to the cap at full reserve before lowering reserve costs 19.98 (5.3 times).
Observed storms follow the same order. There are 922 episodes peaking above $1,000 with a clean two-hour run-up (New South Wales 244, Queensland 296, South Australia 271, Tasmania 72, Victoria 39). Over the 120 minutes before the peak, median reserve falls from 0.70 to 0.40 of its calm level. The median price rises only from $138 to $298 by 20 minutes before the peak, passes $300 in the last 10–15 minutes, and the median peak is $4,167. In the earlier analysis, 61.5% of storms lie closer to the least-action path than to the straight diagonal.
Storms do not follow the path exactly. Their action is a median 2.2 times the least action between their own start and end points (95% interval 2.1–2.4), and at prices of $1,500 and above they hold a median 0.30 more normalised reserve than the least-action path does. The final step to the peak is a jump, which a drift-and-diffusion description does not represent.
Interpretation. In the 5-minute data the price is not bid to the cap while reserve is ample; the least-action route to the cap passes through low reserve. A falling reserve margin over the preceding one to two hours is the common precursor of major spikes, while the last step gives little warning in the price series: in FY26, the first interval at or above $1,000 was already at or above $5,000 in 80% of the episodes that reached $5,000 (Note 03). Tracking reserve relative to its normal level therefore adds lead time that price alone does not provide.
Jumps carry 83–99% of 5-minute price variance, and no price level at $300 or above is stable
The Kramers–Moyal method estimates, for each price level, the mean of the next 5-minute change (the drift) and its variance. In every region the drift has a single stable point, where it turns from upward to downward: $30 in New South Wales, $24 in Queensland, $12 in South Australia, $32 in Tasmania and $25 in Victoria. The effective potential built from drift and variance has a single minimum in each region. There is no second stable level at $300 or above.
The variance is not diffusive. Within each price bin, a move larger than five robust standard deviations (based on the median absolute deviation) counts as a jump. Across the populated price range, −$90 to $600, jumps carry a median 85% of 5-minute variance in New South Wales, 87% in Queensland and South Australia, 83% in Victoria and 99% in Tasmania. A Gaussian sample with the same variance puts close to none of its variance beyond that threshold. The exception is a band of low positive prices, about $3–$40, in New South Wales, Queensland and Victoria, where changes are more evenly sized and jumps carry 1–45% of variance; those bins hold about 7% of 5-minute changes in each of the three regions. In South Australia, jumps carry 84% of variance between $20 and $80 (95% interval 83–85%) and 97% between $300 and $1,000.
Most intervals barely move: 8% (South Australia) to 31% (Tasmania) of 5-minute changes are exactly zero, and the median excess kurtosis of 5-minute changes in South Australia's price bins is 27, against 0 for a Gaussian.
Interpretation. The price is a mean-reverting level with intermittent jumps: most of the time it holds or moves slightly, and at most price levels most of its variance comes from a minority of large moves. Models that represent price volatility as continuous diffusion, with Gaussian changes and a level-dependent variance, misallocate that variance. The quantities that matter for forecasting are the rate and size distribution of jumps at each level. The $300 threshold used throughout this series to define spikes is not a stable state of the price dynamics; it is a convention for measuring them.
14% of prices at or above $5,000 sit exactly on the cap, a point mass that continuous tail models miss
The price has two administered limits: the market floor at −$1,000 and the market price cap, which rose from $15,100 in FY22 to $20,300 in FY26. A visit is a run of consecutive valid intervals within $1 of a limit. There were 148 visits to the cap and 88 to the floor. The longest visit lasted 24 intervals (two hours) at each limit. The mean visit lasted 1.89 intervals at the cap and 1.84 at the floor, and 29% and 25% of visits lasted two intervals or more. A reflecting limit with the volatility measured beside it would return the price to the interior within one interval in more than 99.9% of visits.
The two limits hold different shares of the tails. 14.0% of valid prices at or above $5,000 (279 of 1,986) sit exactly on the cap, rising to 38% of prices at or above $14,000. At the floor, 2.1% of prices at or below −$100 (162 of 7,794) sit on the floor, but 48% of prices at or below −$800 (162 of 336) do.
Interpretation. Both limits are sticky: 29% of visits to the cap and 25% of visits to the floor last two intervals or more, whereas a continuous price process reflecting off the limit would almost never stay a second interval. For statistical models of price tails this matters directly. Because 14% of the ≥$5,000 tail is a point mass at the cap, fitting a continuous extreme-value distribution to these prices overstates the probability of continuous exceedances by about 16% (14 ÷ 86). The negative tail rarely reaches the floor, but the prices that reach the deepest part of it are mostly floor prices.
After a disturbance, price oscillates with a period of about 15 minutes and a damping ratio of 0.47, in every region and year
An earlier analysis removed each region-year's time-of-day profile (separately for weekdays and weekends) from the asinh price, removed the slow first-order component (autoregressive coefficient 0.87–0.97), and fitted a second-order autoregressive model to what remained. In all 25 region-years the model has a pair of complex roots: a damped oscillation with a period of 12.4–15.9 minutes (median 15.1) and a damping ratio of 0.41–0.56 (median 0.47). The root modulus is about 0.34 per interval, so most of an oscillation's amplitude is gone within one period.
The power spectrum shows the same feature. Relative to a first-order model with the same total power, the observed spectrum is 2.3–2.6 times higher at periods of 11–16 minutes, where the first-order model's 95% band lies between 0.95 and 1.12; the excess fades by about 40 minutes.
A separate analysis of how price responds to unscheduled net-load changes (demand minus wind, net of their usual daily and weekly pattern) found the same time scale. Per gigawatt of disturbance, the response is largest for disturbances with periods of 10–16 minutes in every region, and largest of all in South Australia: 4.0 asinh units per GW at 10.2 minutes. In that band the linear relation explains only 2–9% of price variance, so the gain is a lower bound on co-movement, not an established causal amplification.
Interpretation. Net of its slow drift, price follows a damped second-order response. In the fitted model, a 5-minute disturbance is followed by a reversal of about a third of it in the next interval (first autoregressive coefficient −0.28 to −0.40), and the small remaining oscillation, with its period of about 15 minutes, is gone within one period. A 5-minute price change therefore carries information about the next interval in particular. This is consistent with fast corrective responses such as rebidding and the dispatch of flexible plant, which the analysis does not separate. Per megawatt, price co-moves most with net-load changes on this time scale, although in that band those changes explain only 2–9% of price variance.
Summary of measured geometry
Before a spike, median reserve falls from 0.70 to 0.40 of its calm level over two hours while price stays below $300 until the last 10–15 minutes; a price-first route costs 5.3 times the least action. On ordinary days price turns first, 5–60 minutes before reserve.
Prices above $300 come in 32% fewer, longer runs than the same prices rearranged with the same spectrum: 33 minutes per run against 22.
Price reverts to one calm level per region ($12–$32), and jumps carry a median 83–99% of 5-minute variance. The cap and floor are sticky (1.8–1.9 intervals per visit); 14% of ≥$5,000 prices sit on the cap.
After a disturbance, price oscillates with a period of about 15 minutes (damping ratio 0.47) in all 25 region-years; about a third of a 5-minute move reverses in the next interval.