Over a backtest spanning January 2020 through September 2025 on the German-Luxembourg day-ahead electricity market, the forecasting function of our Continuous Learning System delivered approximately 14% higher cumulative trading P&L than TabPFN, a transformer-based foundation architecture for tabular prediction published in Nature, and outperformed every classical baseline we tested. The CLS entered the backtest untrained and learned fully online. This report documents the results, the methodology and its limitations.
- Market
- DE-LU day-ahead (SDAC)
- Period
- Jan 2020 – Sep 2025
- Trading days
- 2,097
- Data source
- ENTSO-E Transparency
- Trading strategy
- Fixed, identical for all methods
Cumulative trading P&L earned by each forecasting system in the identical battery trading simulation over the full backtest. Perfect foresight trades on realized prices and marks the theoretical maximum; no real system can exceed it. CLS figure: latest run. Time-series baselines were benchmarked against an earlier CLS run, see table notes.
Key results at a glance
| System | Type | P&L ratio (CLS / benchmark) | Throughput ratio | Forecast MAE ratio3 |
|---|---|---|---|---|
| Perfect foresight2 | Theoretical ceiling using realized prices as the forecast | ~0.58xceiling | ~0.53xceiling | n/a |
| Wakeline CLS (reference) | Continuously learning architecture; entered the backtest untrained | 1.00xreference | 1.00xreference | 1.00xreference |
| TabPFN 2.51 | Transformer-based foundation architecture for tabular prediction (Nature, 2025) | ~1.14xCLS ahead | ~1.00xparity | ~1.20–1.25xCLS behind |
| Prophet2 | Statistical time-series decomposition (Meta) | ~1.10xCLS ahead | ~1.40xCLS ahead | ~0.88xCLS ahead |
| Previous-day persistence2 | Naive baseline: tomorrow equals today | ~1.14xCLS ahead | ~1.15xCLS ahead | ~1.00xparity |
| NeuralProphet2 | Deep-learning time-series architecture with autoregression | ~1.30xCLS ahead | ~1.40xCLS ahead | ~0.85xCLS ahead |
1 Run against the latest CLS iteration (≈ €13.3M vs ≈ €11.7M cumulative P&L). TabPFN was given the same multi-signal feature set the CLS uses, plus a later daily forecast cutoff. Both choices favor the benchmark.
2 Run against an earlier CLS iteration (≈ €11.7M cumulative P&L). The time-series baselines were evaluated with price-only inputs, since these architectures require exogenous variables to be known over the full forecast horizon in advance, which is structurally unavailable in the day-ahead setting. Perfect foresight is a theoretical ceiling, in place of a competitive baseline: the CLS captured roughly 50 to 60% of the maximum arbitrage value available in the market.
3 MAE, mean absolute error: the average absolute difference between forecast and realized price across all hours. The ratio divides the CLS's error by the benchmark's, so values below 1.00 mean the CLS forecasts were more accurate pointwise and values above 1.00 less accurate. Because the columns read in opposite directions (P&L and throughput above 1.00 favor the CLS, MAE below 1.00 does), each cell carries a plain verdict: whether the CLS is ahead, at parity or behind on that measure. Rows are ordered by cumulative P&L, matching the chart above. Why lower pointwise accuracy can coexist with higher trading P&L is explained in the next section.
What this benchmark measures, and what it does not
The Continuous Learning System (CLS) is our AI architecture built around exactly this capability: it enters a deployment untrained and adapts online from live data throughout operation. Forecasting is one function of its continuous perception layer, and this benchmark evaluates that single function through a monetization proxy. The same function runs in production as Market Edge, our forecasting application for European battery storage operators.
- It measures the quality of 24-hour day-ahead price forecasts, expressed as realized trading P&L, throughput and forecast error.
- It compares every forecasting method through the identical battery trading simulation, with identical constraints, costs and bidding logic.
- The trading strategy does not belong to the CLS. It is a deliberately simple, fixed heuristic inside a simulator we built for evaluation purposes. The CLS produces forecasts only and has no influence over position sizing, execution rules or trading logic.
- These results are internal and have not been independently verified. The full setup is documented on this page, and we share the detailed benchmark reports in direct conversations so the methodology can be assessed in depth.
Why day-ahead electricity prices are a hard problem worth solving
Every day at noon, the European Single Day-Ahead Coupling clears one of the largest auctions in the world. Producers and consumers of electricity across the continent submit bids for every hour of the following day, and the resulting prices coordinate which power plants run, when industry consumes and how storage assets operate. On the German-Luxembourg bidding zone alone, this market moves billions of euros per year.
Anticipating these prices has become dramatically harder. Wind and solar now set the marginal price in a growing share of hours, and their output shifts with the weather rather than with a schedule. Days with abundant renewable generation push prices toward zero or below; scarce hours produce sharp spikes. Between these extremes, the market has lived through structural shocks that no historical dataset prepares a forecasting system for: pandemic-era demand collapse, the 2022 energy crisis with prices an order of magnitude above their historical range, and the subsequent normalization. Each regime rewrote the statistical rules of the market within weeks.
This non-stationarity is precisely why we chose the market as a proving ground. A system that is trained once on historical data ages the moment the regime shifts. A system that learns continuously can follow the shift as it happens. Day-ahead prices offer a rare combination for evaluating this claim: a daily, unambiguous ground truth, public high-quality data through the ENTSO-E Transparency Platform, and a direct way to express forecast quality in euros through storage arbitrage.
The economic relevance is growing with the grid itself. Battery energy storage is expanding rapidly across Europe, and every one of these assets earns its revenue by charging in cheap hours and discharging in expensive ones. The value of a storage asset is, to a first approximation, the quality of the price forecast behind it.
Why we measure forecasts in euros
Forecasting research usually reports pointwise error metrics such as mean absolute error. For decision-making systems this is a weak proxy. A forecast that misses the absolute price level by a wide margin can still identify the cheapest and most expensive hours of the day correctly, and that identification is what a downstream decision actually uses.
Our benchmark therefore monetizes forecast quality: every method's 24-hour forecast feeds the same battery trading simulation, which charges in forecasted low-price hours and discharges in forecasted high-price hours under realistic physical and market constraints. Revenue differences between methods stem from forecast quality alone.
The explanation is structural. The CLS captures the action-relevant shape of the price curve: intra-day spreads, ramps and the timing of extremes. This effect is strongest during volatile regimes such as the 2022 energy crisis, precisely the periods that dominate realized economic outcomes. A method can win on average error and lose where it matters.
The simulated battery has 200 MWh capacity and 100 MW charge and discharge power, with 90% round-trip efficiency, €1/MWh degradation cost and €0.20/MWh market fees. Bids only execute when they clear against actual SDAC prices, so forecast errors translate into missed trades exactly as they would in the real market. There is no intraday rebalancing and no access to future information.
What the CLS was tested against
The suite spans fundamentally different forecasting paradigms, from a zero-parameter reality check to a pretrained foundation architecture.
TabPFN 2.5 is a transformer-based system for tabular prediction, published in Nature in January 2025 and widely regarded as state of the art. We implemented it as 24 independent regressors, one per delivery hour, retrained daily on a rolling 60-day window with the same multi-signal inputs the CLS uses: load, wind, solar, prices and cross-border flows.
NeuralProphet is a PyTorch-based extension of Prophet with a learned autoregressive component over the last 168 hours of prices. It was the most computationally expensive baseline in the suite, retrained daily.
Prophet is Meta's additive decomposition of trend, seasonality and holiday effects: an interpretable, production-grade standard that is widely used in the energy sector, retrained daily on a 60-day window.
Naive persistence and perfect foresight bound the problem. The lower bound assumes tomorrow's prices equal today's, a surprisingly strong baseline in electricity markets. The upper bound trades on realized prices as if they were known in advance and defines the total arbitrage opportunity in the market.
Where the setup favors the benchmarks
Several design choices bias the comparison against the CLS. We document them because they make the results conservative.
TabPFN received a later forecast cutoff. The TabPFN forecast is executed at 14:00 local time, while the CLS forecast is retrieved around 11:00, before gate closure at 12:00. TabPFN therefore sees up to three additional hours of load, renewables and cross-border updates on every trading day.
TabPFN received the full multi-signal feature set. The time-series baselines were evaluated price-only for methodological cleanliness. TabPFN, which consumes decision-time snapshots by design, was given the same system-level signals the CLS uses.
NeuralProphet received post-gate price data. Due to its autoregressive input window, NeuralProphet was fed the complete current-day price curve, beyond the CLS gate-closure cutoff, because this materially improved its performance.
The CLS entered every run untrained. The CLS starts with zero pretraining and learns fully online during the backtest, with an expected stabilization phase at the beginning. TabPFN carries extensive offline pretraining on millions of synthetic datasets. Accumulated metrics start after a 10-day warm-up so that all methods have a minimally valid training window; this applies identically to every run.
A continuously learning architecture that measurably improves
Because the benchmark setup is frozen, we can rerun it after internal development iterations and attribute every change in outcome to the architecture itself. Between two recent iterations of the perception layer's synthesizer, the component that constructs full 24-hour price curves from learned internal representations, cumulative P&L against TabPFN moved from parity to approximately 1.14x. Input data, decision timing, simulator logic and evaluation metrics were unchanged.
The improvement was regime-selective. Error reductions concentrated in periods of extreme volatility, late 2021 through early 2023, where mistakes are most costly and opportunities are largest. Throughput stayed at parity, which means the gain came from better forecasts under a fixed strategy instead of higher trading intensity or added risk. Relative performance remained stable across the full horizon, with no late-stage drift that would indicate transient tuning.
We consider this the most important finding on this page. It shows the architecture improving along economically meaningful dimensions under unchanged, conservative evaluation conditions, which is the behavior a continuously learning system is designed for.
Known limitations
The CLS maintains a persistent pointwise accuracy gap relative to TabPFN, with an average MAE ratio around 1.2x. This does not prevent superior trading outcomes in this setting, and reducing the gap remains an active objective for applications where absolute accuracy matters more directly.
The CLS advantage is strongest in stress regimes and closer to parity in calm markets. The early portion of every run shows elevated variance while the system stabilizes, which is expected for an online learning setup and is excluded from accumulated metrics via the warm-up rule described above.
Finally, these are backtest results on historical data from one market. They quantify forecast quality under realistic market-clearing conditions; they are no guarantee of live trading performance, and we do not offer them as such.
Frequently asked questions
What evidence is there that Wakeline's technology works?
This page is the evidence we publish: a 2,097-day backtest on the German-Luxembourg day-ahead market in which our continuously learning architecture outperformed TabPFN, Prophet, NeuralProphet and a naive persistence baseline on realized trading P&L, with the full methodology and its limitations documented alongside. The same forecasting function runs in production in Market Edge, our application for European battery storage operators. The results are internal and have not been independently verified; we welcome replication and supervised evaluations with partners.
Were these results independently verified?
No. All results are internal and were produced by us. We document the full benchmark setup, data sources, simulator constraints and known caveats on this page so the methodology can be assessed directly. We welcome independent replication and are open to supervised evaluations with partners.
Is the trading strategy part of the CLS?
No. The CLS produces forecasts only. Trading decisions are made by a separate battery trading simulator that we built for evaluation purposes. The same simulator, with the same fixed strategy, processes the forecasts of every method in the comparison, so revenue differences isolate forecast quality.
Why is the CLS forecast error higher than TabPFN's while its P&L is higher too?
Average pointwise error rewards being close to the price level everywhere. Trading value rewards getting the shape of the day right: which hours are cheapest, which are most expensive, and how large the spread is. The CLS captures this action-relevant structure well, particularly in volatile regimes, which is where most of the economic value in this market is concentrated.
Which market and period were tested?
The German-Luxembourg (DE-LU) day-ahead market within the European Single Day-Ahead Coupling, from January 1, 2020 to September 29, 2025: 2,097 trading days at hourly granularity, using ENTSO-E Transparency Platform data. The period covers several distinct regimes, including COVID-era volatility and the 2022 energy crisis.
Does the benchmark cover the whole CLS?
No. It evaluates one function of the continuous perception layer: forming 24-hour price forecasts from live market state. The broader architecture addresses continuous state maintenance and resource allocation as well; those capabilities are outside the scope of this benchmark.
Can we see the full benchmark reports?
Yes, in direct conversation. The full reports contain simulator mechanics and per-benchmark implementation details, and we prefer to walk technical readers through them personally. Implementation-level details of the CLS itself are withheld pending patent proceedings. Reach out via our contact page and we will set up a session.
Want to look under the hood?
This page summarizes the results. The full benchmark reports contain the complete simulator mechanics, per-benchmark implementation details and the diagnostic plots behind every figure shown here. We walk interested partners and journalists through them directly, because the interesting conversations start where the summary ends. Talk to us about the benchmarks.
Disclaimer: All results on this page were produced internally by Wakeline GmbH and have not been independently verified. Figures are approximate cumulative values over the stated backtest period and are rounded. The battery trading simulation exists to monetize forecast quality under realistic constraints; it does not represent a live trading system, and nothing on this page constitutes a forecast of live performance or an investment recommendation. Prophet and NeuralProphet are open-source projects of their respective maintainers; TabPFN is published by Prior Labs. We report their results under the configurations documented in our benchmark reports, which we share in direct conversations.