A backtesting framework for a mean-reversion strategy on European cross-border day-ahead power spreads. The strategy itself is deliberately simple; the point of the project is the evaluation around it, which is built to answer two questions honestly: is the statistical edge real, and is it actually realizable after the frictions a price-series backtest cannot see.
The short version: the strategy finds a strong, well-identified mean-reversion signal that produces an out-of-sample Sharpe above 5 and survives transaction costs up to about 17 EUR/MWh. It also explains, with evidence, why that Sharpe is not a live trading edge. A backtest that prints a Sharpe of 5 and stops there is usually hiding something; most of this repository is the work of finding out what.
Data is real and free: hourly day-ahead prices for Germany and six neighbouring zones from SMARD.de (the Bundesnetzagentur's platform), 2021–2024. No API key.
Project developed May 2026.
Seven instruments (the German price level plus six DE-vs-neighbour spreads) traded as a daily, equal-weight, mean-reversion portfolio. The sample is split 50/50 into in-sample and out-of-sample, and all headline figures are out-of-sample and net of a 0.50 EUR/MWh round-trip cost:
| Metric (out-of-sample, net of costs) | Value |
|---|---|
| Annualized Sharpe | 5.45 |
| Sortino | 8.01 |
| Max drawdown | −101 EUR / MWh notional |
| Calmar | 9.9 |
| Hit rate | 62.5% |
| Break-even round-trip cost | ≈16.6 EUR/MWh |
A Sharpe of 5 on a daily strategy should be treated as a warning, not a result. Real funds operate at Sharpe 1 to 2. When a backtest prints 5, the useful instinct is to look for the leak (lookahead, survivorship, an untradable instrument) before trusting it. The signal here turns out to be genuine, but the realizable edge is a smaller and more complicated thing.
There is a second point worth noting up front. Tripling the assumed cost from the original 0.15 to 0.50 EUR/MWh barely moved the Sharpe, because break-even sits near 17 EUR/MWh, far above any plausible execution cost. Cost is not what limits this strategy.
If the apparent edge were an accounting leak, delaying execution would not remove it. A genuine one-day mean reversion has to decay as you wait, and it does:
| Execution lag | Gross Sharpe |
|---|---|
| 1 day | 5.09 |
| 2 days | 3.47 |
| 3 days | 1.41 |
| 5 days | −0.40 |
The value lives entirely in the next-day reversion, which is a known and fragile property of power spreads.
Every combination of lookback window and entry threshold is profitable in-sample (Sharpe roughly 2.4 to 6.6). There is no single lucky cell; the result is a broad plateau, which is what a real effect looks like and an overfit one does not.
Shuffle each instrument's daily changes to destroy the time structure while keeping the distribution, then re-run. The edge should vanish, and it does: across 50 surrogates the out-of-sample Sharpe is −0.05 ± 0.62, against 5.45 on the real data, a separation of about 8.9 standard deviations. The profit comes from genuine autocorrelation, not from the machinery of the backtest.
The edge survives costs (break-even near 17 EUR/MWh, roughly twice the typical daily spread move), so transaction costs are not the binding constraint.
What limits the realizable edge is tradability, the part a price-series backtest cannot represent:
- Cross-border spreads are not freely tradable. Capturing a DE-AT spread requires cross-border transmission capacity, which is auctioned (FTRs and PTRs). Much of the spread's value is already priced into those rights rather than left available to a price-taker.
- The instrument is a financial abstraction. The backtest marks a daily position to consecutive day-ahead clearings, which is economically a spread swap or CfD. Those instruments exist but are far less liquid than the underlying and carry a wide bid-ask, and the day-ahead auction itself clears only once, so you cannot round-trip inside it.
- Market impact at size. The result assumes zero impact. Any meaningful volume moves these less-liquid spreads against you.
- Persistence is itself evidence of friction. A freely capturable Sharpe-5 effect would not survive four years. That it persists is a sign the frictions above are real.
Conclusion. There is a real, robust mean-reversion signal in cross-border day-ahead spreads. Turning it into P&L is a transmission-rights and liquidity problem, not a signal problem, and a backtest that ignored that would be exactly the overfit story this project is built to avoid.
Instruments. The reference-zone (DE/LU) daily-baseload price level, plus the spread of DE/LU against each of FR, NL, BE, AT, CH and DK1. Each is traded as a daily-settled financial position (a swap or CfD for the spreads).
Signal. For each instrument, z-score the level against a trailing window and
fade the deviation: short when the z-score is high, long when it is low, flat
near the mean. Entry and exit use a hysteresis band (entry_z=1.5 above
exit_z=0.5) to avoid churning in and out.
No lookahead. The z-score on day t uses prices only up to day t, and the resulting position earns the t to t+1 move, enforced by a one-day shift in the P&L accounting:
gross_pnl[t] = position[t-1] * (level[t] - level[t-1])
cost[t] = |position[t] - position[t-1]| * (fees + slippage)
net_pnl[t] = gross_pnl[t] - cost[t]
Portfolio. Equal weight across the seven instruments. Diversification across imperfectly correlated spreads is what lifts the portfolio Sharpe above the roughly 2.9 to 4.2 of any single instrument.
Sharpe, Sortino, max drawdown, Calmar, hit rate, turnover and annualized P&L,
all computed out-of-sample and net of modeled costs, with the gross (no-cost)
figure alongside so the cost impact is explicit. Everything is written to
results/*.csv for auditing.
See RUNNING.md for full setup, including a VS Code walk-through. The short version:
pip install -r requirements.txt
python scripts/run_all.py # downloads data once, then runs the full analysisIndividual steps:
python scripts/download_prices.py
python scripts/run_backtest.py
python scripts/run_cost_sensitivity.py
python scripts/run_robustness.pyTests:
pytestpower-spread-trader/
├── config/strategy.yaml # zones, signal params, costs, OOS split (no magic numbers in code)
├── src/spreadtrader/
│ ├── data/smard_prices.py # SMARD multi-zone day-ahead price client (cached)
│ ├── config.py # typed config and validation
│ ├── signals.py # instruments and z-score mean reversion (hysteresis, no lookahead)
│ ├── backtest.py # engine: positions to P&L, costs, OOS split
│ ├── metrics.py # Sharpe, Sortino, drawdown, Calmar, hit rate, turnover
│ └── plots.py # charts
├── scripts/ # download, backtest, cost sensitivity, robustness, run_all
├── tests/ # no-lookahead, hysteresis, cost accounting, null and trend sanity
├── notebooks/ # exploratory scripts
└── results/ # generated charts and metrics CSVs
| Field | Zones | Source |
|---|---|---|
| Day-ahead price (daily baseload) | DE/LU, FR, NL, BE, AT, CH, DK1 | EPEX SPOT / Nord Pool via SMARD |
- SMARD.de (Bundesnetzagentur), open data portal: https://www.smard.de/en
- Market data download centre and terms of use: https://www.smard.de/en/downloadcenter/download-market-data/
- Day-ahead prices are set by market coupling across EPEX SPOT and Nord Pool bidding zones.
Hourly prices are fetched in weekly chunks, cached to data/raw, and averaged to
a daily baseload product. Timestamps are parsed as UTC and converted to
Europe/Berlin so daylight-saving days are handled correctly.
- Price-taker, financial settlement. Positions settle cash against the day-ahead clearing, with no transmission-rights modelling and no order book. This is the main reason the backtest Sharpe overstates the realizable edge, as discussed above.
- Daily baseload. Hourly prices are averaged to a daily baseload product to give a clean position strategy; hourly or intraday execution is an extension.
- No market impact, constant per-unit costs, perfect fills.
- The signal is intentionally simple. The contribution here is the honest evaluation, not signal sophistication.
- Hourly or intraday execution against continuous intraday prices, where a genuine day-ahead-to-intraday round-trip exists.
- Transmission-rights-aware P&L, netting the FTR/PTR auction cost against the spread capture to estimate the accessible edge.
- Volatility-scaled position sizing and a portfolio risk budget.
- Walk-forward re-fitting of the lookback and threshold rather than fixed parameters.
Mohammad Faisal, M.Sc. Power Engineering (Renewable Energy)
- GitHub: github.com/CodingxFaisal
- Email: mohammad.faisal@gmx.de
Released under the MIT License. See LICENSE.
Data belongs to SMARD.de / Bundesnetzagentur and is used under their terms; see the download-centre link above.