Auction backtest · BNetzA Wind an Land

Model bid band vs. observed award prices, 2017–2025

For every onshore auction round, the engine solves each sampled farm's break-even bid at a 7 % equity hurdle — then the model band is laid over what BNetzA actually cleared. The claim is falsifiable by construction.

Method

Does the engine's break-even bid reproduce observed BNetzA onshore clearing prices 2017–2025? Four steps, run identically for every round:

  1. 1

    Sample the round's pipeline

    For each auction year Y, sample N=40 farms from MaStR commissioned in Y..Y+2 — a proxy for that round's project pipeline — stratified by Bundesland wind class (north / center / south).

  2. 2

    Vintage the assumptions

    Assumptions per year come from cost_vintages.csv (capex, opex, interest-rate vintage). Wind resource is estimated via Path B for every farm — a uniform method across years, because comparability beats precision here.

  3. 3

    Solve the break-even bid

    For each sampled farm, the engine solves the break-even anzulegender Wert (§8.6) such that equity IRR equals the 7 % target — producing a distribution of model bids per round.

  4. 4

    Compare against reality

    The model's median and P25–P75 band are compared with the actual average award price (Ø-Zuschlagswert) per round. Reported: MAE in ct/kWh and the directional hit-rate — did the model move the right way round-over-round.

Caveats

Five reasons to distrust the chart above — stated up front, because the honesty is the feature.

Winner's curse

Auction winners are systematically the most optimistic bidders. Observed awards therefore sit below an unbiased break-even estimate; the model prices a median project, not the most aggressive bid in the room.

Undersubscribed rounds 2019–2022

Many rounds in 2019–2022 were undersubscribed. With little competitive pressure, awards drifted toward the Höchstwert cap — prices in those rounds reflect the cap more than project economics.

Höchstwert binding post-2022

After the 2022–2023 cost inflation, the Höchstwert became a binding constraint: awards cluster at the cap, so the observed series carries regulatory truncation the model does not.

Site-selection bias

Farms sampled from the registry are commissioned projects, not the bid pipeline. Registry ≠ bids: projects that bid and lost, or won and were never built, are invisible to this sampling.

Path B resource error

The backtest uses the coarser Path B wind-resource method uniformly for comparability. Its per-site capacity-factor error propagates directly into each farm's break-even bid.