Post: Forecasting performance matters. Very few providers show the data.

Forecasting performance matters. Very few providers show the data.

Guest post by GridBeyond Business Development Director (AU), Mark Netto

Forecasting is often treated as a “black box” capability in battery optimisation—but it has a direct impact on project returns.

It’s not just about predicting absolute prices correctly. It’s about:

• identifying high-value intervals
• getting the daily price shape right
• avoiding poor charge/discharge sequencing
• navigating increasingly volatile and negative pricing conditions

We compared our forecasts against actual NEM outcomes across all five regions and benchmarked them against Australian Energy Market Operator forecasts.

The results showed material differences—particularly during the intervals that matter most commercially.

As battery margins tighten, this is becoming harder to ignore.

Forecasting is one of the most frequently referenced capabilities in battery optimisation, yet it remains one of the least rigorously assessed. Most providers in this market claim to have superior forecasting capability, often referencing artificial intelligence, machine learning, or proprietary optimisation models. Far fewer explain how forecasting performance is actually measured, or provide the underlying data required to validate those claims.

That gap is becoming increasingly important.

Battery revenues in the NEM are not evenly distributed across time. A meaningful proportion of annual revenue is concentrated in relatively short periods of price volatility, including price spikes, supply shortages, rapid demand shifts and unexpected market reversals. Missing these windows—or being positioned incorrectly during them—can have a disproportionate impact on annual returns.

A forecasting error during low-demand overnight periods may have minimal financial consequences. The same error during a high-price interval can materially impact project performance. This becomes particularly important when considering that intervals above $300/MWh typically represent less than 2% of total dispatch intervals, yet often contribute disproportionately to annual battery revenues. This asymmetry is one of the primary reasons forecasting performance matters far more than average market conditions might suggest.

This is also where performance differences between optimisation platforms tend to become most visible.

Forecasting accuracy is not simply about predicting absolute prices correctly—it is also about getting the shape of the day right.

For battery operators, understanding when volatility occurs can be just as important as understanding how high prices may move. A forecast that predicts a price spike but gets the timing wrong can still lead to poor dispatch outcomes. Charging too early, discharging too early, or failing to preserve state of charge ahead of an evening peak can materially reduce annual returns.

This becomes even more important as batteries increasingly optimise across both energy and FCAS markets.

The benchmark problem

Most forecasting comparisons in the market use Australian Energy Market Operator P5 pre-dispatch forecasts as the benchmark. While this is a useful reference point, it should not be treated as evidence of forecasting capability.

P5 is a publicly available signal accessible to all market participants. Outperforming it should be considered a baseline expectation rather than a differentiator.

A more meaningful assessment considers how forecasting models perform against actual spot price outcomes over a full year across all NEM regions. This typically requires two metrics.

Mean Absolute Error (MAE) measures average forecast error across all intervals and provides a view of normal forecasting performance.

Root Mean Square Error (RMSE) places greater weighting on larger forecast errors, making it more sensitive to extreme pricing events.

Both metrics matter. However, RMSE often becomes particularly important in battery optimisation because extreme price events tend to drive a disproportionate share of annual returns. A model with relatively low MAE but weak RMSE may perform adequately under normal market conditions while underperforming during the intervals that matter most commercially.

What the data shows

We recently completed a full-year analysis of our NEM spot price forecasting performance across all five regions, comparing our forecasts against actual spot prices and benchmarking against AEMO across both MAE and RMSE metrics.

The results across mainland regions were material (lower MAE is better!):

• NSW: $28.6/MWh vs AEMO’s $65.2/MWh (56% improvement)
• QLD: $17.6/MWh vs $29.3/MWh
• SA: $25.9/MWh vs $35.6/MWh
• VIC: $19.1/MWh vs $25.1/MWh

Tasmania remains the one region where AEMO currently holds a marginal advantage on RMSE, although we continue to outperform on MAE and NMAE.

These results are directionally important because they demonstrate that forecasting performance can be measured consistently—and that material performance differences do exist between models.

Why South Australia deserves attention

South Australia is becoming one of the clearest examples of why forecasting capability matters.

Approximately 30% of intervals now experience near-zero or negative pricing conditions, largely driven by high solar penetration during daylight hours. These conditions create increasingly complex operating decisions for sub-5MW hybrid assets, particularly around charging behaviour, clipped solar capture, negative price avoidance and preserving optionality for future dispatch intervals.

During these intervals, our Normalised MAE was 30.5%, compared with AEMO’s 61.0%.

This is not a marginal improvement. It reflects a materially different ability to navigate pricing conditions that are increasingly defining South Australia and are beginning to emerge more frequently in Queensland as renewable penetration continues to increase.

What developers should be asking for

The forecasting discussion in this market has historically been dominated by marketing claims rather than measurable performance outcomes. That is beginning to change.

Developers and investors should expect greater transparency around forecasting performance, including regional MAE and RMSE performance, behaviour during high-price intervals, performance during negative pricing events, and visibility over the methodology used to produce these results.

Our NEM data science team led this analysis, and we are happy to share the underlying methodology with developers or investors who want to examine the data in more detail. That level of transparency should become standard practice across the market.

The broader implication

Forecasting capability is not the only determinant of battery performance, but it underpins almost every operational decision that follows. State of charge management, FCAS participation, arbitrage sequencing and hybrid optimisation decisions are all downstream of a forecast.

A platform that consistently underperforms on forecasting will ultimately underperform in the areas that matter most commercially.

As the NEM becomes more competitive and optimisation margins tighten, this issue will become increasingly difficult to ignore.

This article is part of a broader series examining what separates high-performing sub-5MW battery portfolios. The first article explored the six capabilities developers should be thinking about as the market matures (link here).

If you’re currently evaluating optimisation providers, testing project assumptions, or building revenue models, happy to share the underlying methodology and data.

The next article will focus on hybrid optimisation, and why many battery platforms still struggle when solar and storage begin sharing the same connection point.