Charge low. Sell high. Know when.
Battery plans from public grid data, tested against the free benchmark each market already has.
- Australia →tested · live
- Texas →tested · live record
- New Zealand →tested
- Great Britain →tested · results held
- Every result →all markets, wins and misses
How Voltensor works
- Public grid dataPrices and forecasts the market already publishes: AEMO's forecast and actual prices in Australia; ERCOT's day-ahead and real-time prices in Texas
- Voltensor modelLearns only from earlier data and is refitted as new data arrives
- Price forecastEvery interval of the next trading day
- Battery planWhen to charge and when to discharge, scored against the plan from the market's free benchmark (AEMO's forecast in Australia, the day-ahead price in Texas)
- Your decisionYou or your trader place the bids. Voltensor does not connect to, operate or bid for your battery.
A market goes live only after a published historical test. Delivered today on this page, with a sample plan by email on request. There is no API yet: tell us which format you would use (email, file or API).
Next battery plan
Live since 8 Nov 2026. At the latest daily fingerprint (8 Nov 2026 19:07 UTC, Record 1) the log held 17 on-time forecasts. No completed on-time trading day yet; formal evaluation after 30 trading days. Live record →
The plan as a table
| Region | ChargeAEMO · Voltensor (range) · plan price | DischargeAEMO · Voltensor (range) · plan price |
|---|
Yesterday: forecast vs actual
The first live plan is for trading day 9 Nov 2026. Its result appears here after that day ends at 04:00 AEST on 10 Nov 2026.
Half-hour prices for this region
| Half-hour to | Actual | AEMO forecast | Voltensor model |
|---|
Live days as a table
| Trading day | Perfect foresight | AEMO plan | Voltensor plan | Voltensor − AEMO | Typical error AEMO | Typical error Voltensor | Typical error last week's prices |
|---|
A$ per MW for the day, the five regions averaged; gross energy only, not expected revenue. Typical error is the average absolute difference between the day-before forecast (Voltensor: the model's price) and the actual 30-minute price, in A$/MWh. Last week's prices is a simple forecast for comparison: the average actual price at the same half-hour over the seven days that ended before the plan was made.
How it did in months the model never saw
These results are from the first live model (Record 1). The model in use since 9 November (Record 2) was chosen on this same test, so its numbers here would flatter it; its live record is the clean test.
Monthly numbers
| Month | Test days | Voltensor − AEMO plan A$ per MW | Running total A$ per MW |
|---|
Download every test day by region (CSV) (columns explained): what each plan earned at actual prices (perfect foresight, last week's average, AEMO, Voltensor), A$ per MW for the day, for the day-before, same-morning and re-planned tests.
Numbers and 95% intervals
| Region | Last week's average | AEMO | Voltensor | Voltensor − AEMO A$/MW/yr (95% interval) | Weeks Voltensor led | Verdict |
|---|
Including the three months used to choose the design (from 2 Jan 2026), all five regions together: Voltensor − AEMO +A$1,156 per MW per year (95% interval +A$665 to +A$1,675) over 302 days. The model's expected price used alone would have earned more in April–October (+A$2,751 per MW per year over AEMO, 95% interval +A$1,032 to +A$4,164), but over all 302 days it was behind AEMO on average (−A$3,844; interval −A$14,841 to +A$2,568): in January it played down price spikes that AEMO forecast correctly. That is why Voltensor's plans average the two.
Accuracy over the same months (every hourly AEMO run, every half-hour up to 39 hours ahead; average absolute error of 30-minute prices): AEMO A$53.4/MWh, Voltensor's price model A$18.8 (its middle forecast), Voltensor's plan forecast A$34.8. The plan forecast averages the model's expected price with AEMO's forecast. It is less accurate on average than the model alone, but it keeps AEMO's warnings of real price spikes, which is where much of a battery's earnings are. The model's 10–90% price range contained the actual price 77% of the time (target 80%).
What this test leaves out, and sources
- Frequency services (FCAS), which many Australian batteries also earn from.
- Battery wear (except in the re-planning test), network charges, bid and trading costs, taxes and financing.
- Any battery size other than 1 MW / 2 MWh.
Shares are gross energy arbitrage. Dollar figures are annualised from the test days and are not expected revenue.
AEMO PREDISPATCH runs: NEMweb PredispatchIS archive (every hourly run, 19 Oct 2025 – 1 Nov 2026). Actual 5-minute prices: AEMO aggregated price and demand data (PRICE_AND_DEMAND monthly files, 2023–2026).
One day up close: South Australia, 9 October 2026
One example, not a performance estimate. We picked this day before the current model existed because it had an early-morning spike and a long stretch of negative prices.
| Plan chosen from | Charge | Discharge | Earned at actual prices | Capture rate |
|---|
Data table: the day in half-hours
| Half-hour ending (AEST) | AEMO forecast | Voltensor | Actual, average | Actual, highest 5-min |
|---|
Other decision times
The same April–October 2026 test with plans made at other times: on the morning of each day, and re-planned every 30 minutes.
Different battery rules
Re-planned every 30 minutes from the latest hourly AEMO run, looking up to 40 hours ahead. The battery may cycle more than once a day, pays A$5/MWh for wear, and is paid 30-minute average prices. These shares are not comparable with the once-a-day plans.
Numbers and 95% intervals
| Region | Last week's average | AEMO | Voltensor | Voltensor − AEMO A$/MW/yr (95% interval) | Weeks Voltensor led | Verdict |
|---|
No proven advantage over the day-ahead price yet.
Three historical tests and a live record. The day-ahead market price is the free benchmark every Texas battery already has, and so far no Voltensor plan has beaten it by a clear margin, either fixed the evening before or re-planned during the day.
Live record, since 9 November 2026
Each night before midnight Texas time, Voltensor fixes the next day's plan for the five trading hubs and fingerprints it; the fingerprint is stored off the planning machine before the day starts. The next morning the plan is scored at real-time prices, beside the plan from the day-ahead price and a simple forecast from last week's prices.
The first plan is for operating day 9 November 2026. It is fixed and fingerprinted before midnight Texas time on 8 November, and its result appears here once that day's real-time prices are in.
Test 2: a learned forecast with ERCOT's own forecasts, April–October 2026
A gradient-boosted model (LightGBM) learned the gap between the real-time and day-ahead price from ERCOT's load, wind, solar and outage forecasts posted the evening before, recent real-time prices and the day-ahead ancillary-service prices. The test was written down and fingerprinted before any result existed; December 2025 – March 2026 chose between the model and a blend, then April–October 2026 was scored once.
Not clear: Voltensor's blend earned +US$1,162 per MW per year more than the day-ahead price plan (95% range −US$334 to +US$3,043), ahead in 6 of 7 months.
Typical error, US$/MWh: day-ahead price 9.9, last week's average 14.0, Voltensor's blend 10.1 (forecasts and actual prices both limited to −100..1500 US$/MWh).
The same numbers as a table
| Hub | Day-ahead plan | Last week's average | Voltensor |
|---|---|---|---|
| Hub average | 69% | 74% | 73% |
| Houston | 66% | 66% | 70% |
| North | 72% | 73% | 72% |
| South | 67% | 67% | 70% |
| West | 71% | 71% | 75% |
Test 3: re-planning every hour, April–October 2026
Market players earn by reacting during the day, so this test let the battery re-plan at the start of every hour with the newest real-time prices and ERCOT's hourly forecast updates. Written down before any result; the benchmark was the better of two free strategies on December–March: keeping the day-ahead price plan (it won) or reacting to the newest price.
Not clear: re-planning with Voltensor earned +US$351 per MW per year more than keeping the day-ahead plan (95% range −US$820 to +US$1,337), ahead in 6 of 7 months. Reacting to the newest price alone: −US$979 per MW per year.
The same numbers as a table
| Hub | Day-ahead plan, kept | Newest price, re-planned | Voltensor, re-planned |
|---|---|---|---|
| Hub average | 68% | 64% | 70% |
| Houston | 63% | 63% | 62% |
| North | 70% | 62% | 69% |
| South | 64% | 65% | 65% |
| West | 69% | 68% | 71% |
Test 1: a simple model, February 2025 – October 2026
Numbers and 95% intervals
| Hub | Last week's average | Day-ahead | Voltensor | Voltensor − day-ahead US$/MW/yr (95% interval) | Months Voltensor led | Verdict |
|---|
Flexible loads, what this test leaves out, and sources
- Ancillary services, which Texas batteries also earn from.
- Battery wear, fees, trading costs, taxes and financing.
- Node-level prices: results are for the five trading hubs.
ERCOT public data through its API: real-time settlement point prices (report NP6-905-CD) and day-ahead settlement point prices (NP4-190-CD), January 2025 – October 2026. Days with a daylight-saving change are skipped. The first four weeks train the model only. On 5 December 2025 ERCOT changed how real-time prices are set (real-time co-optimisation with batteries, "RTC+B").
Great Britain: first test done, results held
We tested plans against the day-ahead auction price for July–October 2026. The price data comes from Nord Pool through the system operator's data portal, and we will publish the numbers once the data provider confirms we may.
What we tested
- The benchmark: the N2EX day-ahead auction price, the free benchmark a GB battery owner already has.
- Paid at: Elexon's market index price for each half-hour, as a stand-in for short-term wholesale prices.
- The test: written down and fixed before any result; January–June 2026 for the choices, July–October 2026 scored once.
- The revenue covered: wholesale energy only. GB batteries also earn in the Balancing Mechanism and from frequency services, which this test leaves out.
No proven advantage over the operator's forecast yet.
At four grid points in July–October 2026, Voltensor did not show a statistically clear advantage over the system operator's own price forecast. New Zealand has no traded day-ahead energy price, so, as in Australia, the free benchmark is the operator's forecast.
- Result: Voltensor's plan minus the plan from the operator's forecast (PRSL): +NZ$1.22/MW/day (95% range −NZ$9.80 to +NZ$10.47/MW/day), pooled over 122 trading days and four grid points; Voltensor ahead on 58 days, behind on 58, the same on 6. Gross simulated wholesale-energy replay, not expected revenue.
- Capture of the perfect-foresight ceiling, same battery rules: Voltensor 55.9%, PRSL plan 55.3%.
- Mean absolute price error: Voltensor NZ$22.1/MWh, PRSL NZ$22.7/MWh.
- By grid point, NZ$/MW/day: Otahuhu +NZ$11.39, Haywards −NZ$12.14, Benmore +NZ$0.32, Huntly +NZ$5.32.
- Benchmark: the system operator's price-responsive long schedule (PRSL): the latest run strictly before 14:30 New Zealand local time on the previous day, no more than 60 minutes old, covering all 48 periods. It was chosen over the non-responsive schedule on January–June data.
- Model: a gradient-boosted model of the gap between actual prices and the schedule; a quarter of its correction is applied. This first test uses only PRSL, NRSL, their difference and calendar effects. It does not test recent actual prices, weather, hydro storage or HVDC transfer.
- Test: the benchmark and model were selected on January–June 2026, then frozen before the single July–October confirmation (122 trading days; the 27 September clock change left out).
- Battery: 1 MW / 2 MWh, 85% round trip, one 2-hour charge and one 2-hour discharge a day, paid final prices.
- Source status: fallback and proxy prices are audited separately where the source metadata allows; an unresolved status is not inferred from equal prices.
Data: www.emi.ea.govt.nz provided by the Electricity Authority (New Zealand), licensed under CC BY 4.0 (EMI legal information): final energy prices and the operator's long schedules (PRSL, NRSL) at nodes OTA2201, HAY2201, BEN2201 and HLY2201. Results are Voltensor's; the Electricity Authority has not reviewed or endorsed them.
Every result we have, in one place.
Each test was written down and fingerprinted before its results existed, and each live forecast is fixed before the day it covers, then scored against the prices that actually happened. We publish all of them, including the ones that did not beat the free benchmark.
Gain over the free benchmark, each market in its own unit
Each dot is Voltensor's plan minus the plan from that market's free benchmark; each bar is its 95% range. A bar that crosses zero is not clear. Gross energy earnings of a simulated battery, not expected revenue.
| Market | What | Period | Compared with | Voltensor's plan minus the benchmark plan, per MW per year unless stated | Verdict |
|---|---|---|---|---|---|
| Australia | Live record 2 (the current model) | from 8 November 2026 | AEMO's forecast | first result after trading day 10 November | early: formal review after 30 trading days |
| Australia | Historical test (the first live model) | 2 April – 30 October 2026 | AEMO's forecast | +A$1,043 a year (95% range +A$541 to +A$1,614) | Clear |
| Texas | Live record 1 (test 2's blend) | from 9 November 2026 | day-ahead price | first result after operating day 9 November | early: formal looks at 6 and 12 months |
| Texas | Test 3: re-planning every hour | April – October 2026 | day-ahead plan, kept | +US$351 (95% range −US$820 to +US$1,337) | Not clear |
| Texas | Test 2: learned forecast with ERCOT's forecasts | April – October 2026 | day-ahead price | +US$1,162 (95% range −US$334 to +US$3,043) | Not clear |
| Texas | Test 1: a simple model | 5 February 2025 – 31 October 2026 | day-ahead price | about the same share of hindsight earnings (66–70%) | Not clear |
| New Zealand | Test 1: operator-forecast benchmark, four grid points | July – October 2026 | operator's forecast (PRSL) | +NZ$1.22/MW/day (95% range −NZ$9.80 to +NZ$10.47/MW/day) | Not clear |
| Great Britain | Test 1: day-ahead auction benchmark | July – October 2026 | day-ahead auction price | held until the data provider confirms we may publish | — |
Gross energy earnings of a simulated 1 MW / 2 MWh battery, not expected revenue. The battery makes one cycle a day (at most 2 MWh charged). In every test except Texas test 3 that cycle is one 2-hour charge block, then one 2-hour discharge block; Texas test 3 re-plans every hour and may spread its cycle across the day, with no wear cost. The perfect-foresight earnings each test is compared with are a simulated ceiling, not something a battery owner could earn. "Clear" means the 95% range is entirely above zero. Each market's page gives the details, the typical forecast error and the data sources.
The same tests on one common scale
Gross replay uplift as percentage points of each test's own perfect-foresight gross; not expected revenue, not capture rate, and not a like-for-like economic ranking. Markets, periods, grid points and battery rules differ (Texas test 3 re-plans every hour), so the rows keep the table's order and each shows its own-unit result. Perfect foresight is a simulated ceiling, not something a battery owner could earn.
Model versions
- Australia, 8 November 2026 03:05 UTC: live record 1. The first live model. On its April–October 2026 test it beat plans from AEMO's forecast by a clear margin.
- Australia, 8 November 2026 19:12 UTC: live record 2. The same design trained on every hourly AEMO run since January 2023. It was chosen by a rule written down before the candidates' results existed; record 1 was closed with its final fingerprint.
- Australia, tested and not adopted (8 November): AEMO's reserve outlook, weather forecasts and generators' published bids each gave no clear gain.
- Australia, decision rule (8 November, checked on 2025): a rule that trusts the model when prices look calm was not confirmed for plans made the day before, but did clearly better for plans made on the morning and re-planned during the day. It is a candidate for a future record.
- Texas: test 1 (a simple model), test 2 (a learned forecast with ERCOT's own forecasts) and test 3 (re-planning every hour): none clearly ahead of the day-ahead price. Live record 1 from 9 November.
- New Zealand, test 1 (8 November): not clear against the operator's forecast. This first test uses only the operator's two long schedules (PRSL, NRSL), their difference and calendar effects.
- Great Britain, test 1 (8 November): done; numbers held until the data provider confirms we may publish them.
Each new model version starts a new record, fixed in advance; earlier records stay frozen and visible. No version is re-scored or replaced after its results are known.
Live record: forecasts logged before the prices are known
Voltensor logs a forecast of each hourly AEMO run before any of the prices it forecasts are known. Only forecasts made within 60 minutes of their AEMO run count toward the record. No completed on-time trading day yet; formal evaluation after 30 trading days. Backfilled runs are scored separately.
Daily fingerprints
Every entry in the log contains the fingerprint of the entry before it, so changing or deleting any entry changes every later fingerprint. Once a day we publish here one fingerprint that covers the whole log so far.
| Day (AEST) | Forecasts on time | Forecasts late | Plans | Fingerprint (SHA-256) |
|---|
Latest fingerprint made 8 Nov 2026 19:07 UTC. Each day's fingerprint is added to this page after it is made.
The rules of Record 2, the current record, were fixed before its first forecast (8 November 2026, 15:30 UTC). Fingerprint of its rules file: 35aa8a971c51f4ae399c9c527dfc459729ae6dc2922ed9df66036263355536cc. Record 1 (model live_202611, rules fixed 8 November 2026, 03:05 UTC) closed on 2026-11-08 19:12 UTC: its model was replaced by the version chosen on 8 November with the user's approval. Its rules file: ab2c146d4226a3baa297d88a150f18874fcca9b0cc7ea0745f460305535fa814; its last daily fingerprint: b629a6ecd4c70b45137c876fa39d66aeab01b4963b0d655d87535daec1201684.
What the fingerprint covers, and how to check it
Each fingerprint is the SHA-256 of a small record: the fingerprints of the latest forecast entry and the latest plan entry, the number of entries, the model files in use and the previous fingerprint. Runs from 1 November were forecast after the fact to start each record's log; they are marked as late and are scored separately. When we share the log, anyone can recompute every fingerprint and compare it with the ones published here. Until then, this page publishes a hash and a timestamp only: readers cannot independently verify the log's contents or timing.
Every published fingerprint in full:
How we test
A historical replay, not a return forecast: in every market, the same battery, the same rules and only the information available at each decision time.
The same rules in every market
The battery
1 MW / 2 MWh with 85% round-trip efficiency. Each day it charges for one two-hour block and discharges for one later two-hour block, and is paid actual prices. 100%, the perfect-foresight ceiling, is the best pair of blocks, known only in hindsight. (Australia's re-planning test uses different rules; see below.)
What we compare
- Last week's average: the price in each interval of the day over the seven most recent days known at decision time, averaged.
- The market's free benchmark: what every battery owner there already has: AEMO's PREDISPATCH forecast in Australia, the day-ahead market price in Texas.
- Voltensor: our forecast, made only from data available at decision time.
Keeping the test honest
Every forecast uses only data published before its decision time, and every model is fitted only on earlier data. Each market's test period and design choices are listed below.
How sure we are
We resample the test days 3,000–4,000 times and report the 95% interval of the daily difference between two plans, annualised. A result counts as clear only when the whole interval is on one side of zero. We also count the weeks or months each plan was ahead.
Going live
A market goes live only after its historical test is published. Live forecasts are logged and fingerprinted before the prices they forecast are known, and forecasts made late don't count.
Market by market
Australia (NEM)
- Benchmark: AEMO's PREDISPATCH forecast.
- Model: The model in use since 9 November 2026 (Record 2) is trained on every hourly AEMO PREDISPATCH run since January 2023 (AEMO's monthly history files and the weekly archive), with price patterns taken from 2019-2022 prices; its inputs and settings are otherwise those below. The April–October test below was made by Record 1's model, a gradient-boosted model (LightGBM) trained on every hourly AEMO PREDISPATCH run since October 2025: forecast prices in all five regions, demand, available supply, wind and solar, batteries, interconnector flows and limits, and binding network constraints, plus how AEMO's forecast changed between runs, recent actual prices and price patterns since 2023. For every half-hour up to 39 hours ahead it forecasts a price range, the chance of a spike above A$300 and an expected price. It is refitted every month on earlier data only. For battery plans, Voltensor averages the model's expected price with AEMO's own forecast.
- Test: We used January–March 2026 to choose the design, including averaging with AEMO's forecast, then fixed it and scored April–October 2026 once. Each monthly refit sets its own model size from its training data only. Record 2's design was then chosen on that same April-October test, by a rule written down before the candidates' results existed: longer history beat versions with AEMO's reserve outlook or weather forecasts. On that test it scored +A$1,318 per MW per year on day-before plans, a number that flatters it because the choice used it; its live record from 9 November is the clean test.
- Decision times: the day before at 13:00 AEST, from AEMO's run labelled 13:00 (published about 12:30) and actual prices up to 13:00; the same morning at 04:00 AEST, from the run labelled 04:00 (published about 03:30) and actual prices up to 04:00.
- Re-planning test: re-planned every 30 minutes from the latest hourly AEMO run, looking up to 40 hours ahead, with different battery rules: it may cycle more than once a day, pays A$5/MWh for wear and is paid 30-minute average prices.
- Live timing: forecasts are scheduled at 7 minutes past each hour from AEMO's newest run, typically about 35 minutes after AEMO publishes it; a forecast made more than 60 minutes after its run is logged as late and does not count. The day-before plan is scheduled to appear on this site at about 13:10 AEST; if an update fails, the page keeps the previous plan and flags it as out of date after 30 hours.
Texas (ERCOT)
- Benchmark: the day-ahead market price.
- Model: an expanding-window ridge regression refitted every day on earlier days only, from the day-ahead price, last week's average, the most recent complete day, time of day and a weekend flag (prices capped at −100 to 1,500 US$/MWh for fitting).
- Test: 5 February 2025 – 31 October 2026, after four weeks used only for training; days with a daylight-saving change are skipped.
- Decision time: midnight Central Time at the start of each operating day, knowing that day's day-ahead prices and real-time prices up to the day before.
- Later tests and the live record: on the Texas page.
Great Britain
- Benchmark: the N2EX day-ahead auction price.
- Numbers are published once the data provider confirms we may: details.
New Zealand
- Benchmark: the system operator's price-responsive long schedule (no day-ahead market, as in Australia).
- Result and method on the New Zealand page.
Who's behind Voltensor
Voltensor applies AI to power markets. It is independent, founder-led and based in Hong Kong and Australia.
Built to be checked, not taken on trust
Voltensor is new and small, so every claim on this site comes with a way to check it.
- Against the free benchmark: every result is compared with what battery owners in that market already get for free, on days the model never saw.
- Forecasts on the record: where Voltensor is live, every forecast is logged and fingerprinted before the prices are known, and late ones don't count. See the live record →
- Open numbers: each result shows what it leaves out, and the live market's daily results, from its test and its live record, can be downloaded. Download the test days (CSV)
Voltensor is not affiliated with or endorsed by AEMO, ERCOT or any other market operator; it uses their public data. Contact: hello@voltensor.ai.
What's next
More markets, each going live only after its historical test is published.
Join the early-access list
Get a sample plan for your market and updates as Voltensor opens to customers. Email and market are enough; battery details are optional.
Email hello@voltensor.ai with your market (Australia, Texas or Great Britain) and, if you like, your battery's size. We'll send a sample plan and tell you when Voltensor opens to customers in your market.
We use your email only to reply and to send the Voltensor updates you asked for. Ask us at any time to delete it.
Prefer to check the numbers first? Download every Australian test day (CSV), no sign-up needed.