CastValue — Quantifying the EUR/MW Value of Forecasts in Energy Trading

CastValue is an open, multi-benchmark framework that measures what a forecast is actually worth in a real trading/dispatch decision — in euros, not error metrics.

This repository currently contains one fully built benchmark (castvalue.bess.daa, described below) plus a second one under active development (castvalue.pv.imbalance, a PV imbalance-cost benchmark) and a placeholder for a third (castvalue.bess.multimarket, scope only — see src/castvalue/bess/multimarket/README.md). The rest of this README describes the original, fully-built BESS day-ahead-only benchmark.

Most forecast providers sell on MAE. Most traders buy on MAE. CastValue shows why that is the wrong metric, and replaces it with one that directly reflects trading P&L.

The Problem

A forecast with a low MAE can destroy more trading value than one with a high MAE. This is not a theoretical edge case — it happens regularly, and this repository demonstrates it empirically on real German DAA price data.

The reason: what matters for BESS dispatch is not how close the forecast price is to the actual price, but whether the forecast correctly ranks the cheap and expensive hours of the day. A forecast can be numerically accurate yet systematically wrong about timing — causing the optimizer to charge when it should discharge, and vice versa.

What CastValue Measures

CastValue is defined as the annual DAA trading margin generated by optimizing BESS dispatch on a given forecast, minus the margin generated by a naive persistence baseline (yesterday's prices as today's forecast):

CastValue(k) = AnnualMargin(forecast k) − AnnualMargin(persistence)

A positive CastValue means the forecast adds real EUR/MW above what any trader could achieve with zero forecasting effort. A negative CastValue means the forecast is worse than doing nothing.

Why This Matters

For BESS traders and asset managers: Stop evaluating forecast providers on MAE. CastValue tells you directly how much each forecast contributes to your annual trading margin, in the currency that appears on your P&L. Use this tool to compare your current provider against alternatives — and against the theoretical ceiling of perfect foresight.

For forecast providers: Benchmark and showcase your forecast in the metric your clients care about. A CastValue number is a commercially credible claim: it tells potential customers exactly how many EUR/MW per year your forecast is worth, backed by a transparent, reproducible methodology.

The MAE Paradox

The backtester surfaces a systematic finding we call the MAE Paradox: days with high forecast error (high MAE) do not reliably lead to high profit loss, and days with low forecast error do not reliably lead to high profit capture.

We classify days into two categories:

• Bad Forecast, Good Capture — high MAE, but the optimizer still dispatched at the right times
• Good Forecast, Bad Capture — low MAE, but a subtle timing shift caused the optimizer to miss the most profitable spreads

This has direct implications for how forecast quality should be specified in commercial contracts.

Methodology

The backtester runs a linear programme (LP) for each day of the dataset, optimising BESS charge/discharge decisions under realistic physical and market frictions:

• Round-Trip Efficiency (85%) in the LP energy balance — every MWh delivered to the grid requires ~1.176 MWh purchased, so the LP only schedules trades that survive RTE losses.
• Opportunity-cost bidding — charge bids = avg forecast discharge price × RTE; discharge asks = avg forecast charge price ÷ RTE. These are the realistic limit prices a desk would post given the forecast.
• Out-of-the-money rejection — if the actual clearing price is unfavourable vs. the posted bid/ask, the MTU does not fill.
• Volume-balanced paired rejection — when OTM fills leave charge and discharge volumes unbalanced, worst-margin MTUs are dropped on the oversubscribed side until volumes match. This forces end-of-day SOC = start-of-day SOC, so forecast accuracy directly drives the fill rate.
• Perfect Foresight is defined as the LP optimum given RTE, with bids placed at clearing prices (100% fill rate). It still pays the RTE loss — a "perfect scheduling" benchmark, not a frictionless one.

Three scenarios are evaluated for each day:

Trading decisions are optimised on forecast prices. Realised margin is always calculated on actual prices.

The dataset covers DE-LU at native 15-minute resolution from 18 October 2025 through 22 February 2026 (126 days).

For the full mathematical formulation see docs/methodology/DAA_DE_LU_Formulation_v1.1.pdf.

Earlier methodology (v1.0): the original single-cycle deterministic dispatch (no RTE, no OTM, fixed daily SOC=0) is preserved at the v1.0 git tag. Check out that tag to access the v1.0 notebook, optimizer, formulation PDF, and whitepaper.

Repository Structure

CastValue/
├── README.md
├── LICENSE
├── DAA_optimizer.py                     # backward-compat shim -> castvalue.bess.daa.optimizer (deprecated)
├── pyproject.toml
├── src/castvalue/
│   ├── core/                            # shared machinery: coalitions, Shapley/Harsanyi, pricing, ceiling checks
│   ├── bess/
│   │   ├── daa/optimizer.py             # this benchmark: Pyomo/GLPK LP optimizer
│   │   └── multimarket/                 # planned, not implemented — raises NotImplementedError
│   └── pv/imbalance/                    # PV imbalance-cost benchmark, under active development
├── examples/
│   └── DAA_DE_LU_Bench.ipynb            # main analysis notebook for this benchmark
├── data/
│   └── master_energy_data_v2.csv        # 15-min DE-LU dataset (Oct 2025 → Feb 2026)
├── docs/
│   ├── methodology/                     # DAA_DE_LU_Formulation_v1.1.pdf/.tex
│   ├── whitepaper/                      # whitepaper PDF
│   └── img/                             # wp_annual_results.png, wp_cumulative_margin.png, wp_example_day.png
└── tests/
    └── bess/daa/golden/                 # golden regression fixtures guarding this benchmark's numerics

The notebook is structured in four sections:

1. Optimisation & backtest — runs the LP for all days and all scenarios
2. Annual results — cumulative margin chart and CastValue bar chart
3. Example day — price forecasts, dispatch decisions, and realised margin for a single day
4. The MAE Paradox — scatter plot and deep-dive on the paradox between forecast accuracy and profit capture

How to Benchmark Your Own Forecast

To run CastValue on your own forecast data, prepare a CSV with the following columns:

Replace the daa_em_forecast column in data/master_energy_data_v2.csv with your own data, update the scenario name in examples/DAA_DE_LU_Bench.ipynb, and run all cells. The notebook expects native 15-minute resolution.

Installing

pip install "castvalue_bench[bess]"

This benchmark also needs a GLPK solver binary on your system (not a pip package) — e.g. conda install -c conda-forge glpk, then point Optimizer(solverpath_exe=...) at the installed glpsol binary. Running the notebook additionally needs pandas, matplotlib, and jupyter.

About

CastValue is developed by FlexUp GmbH. For enquiries about forecast benchmarking or CastValue licensing, please get in touch.