Clipping battery-arbitrage forecasts to the price cap can cut error and still erase profit: 1,158 Indian market days
Synopsis
In battery arbitrage, when a battery starts and ends the day at the same charge, a common shift in prices affects its schedule only through round-trip losses, so decision-focused losses identify the forecast level only weakly; clipping forecasts to a price cap can lower prediction error while removing profitable spreads. On 1,158 holdout days of the Indian day-ahead market, an equal-weight average of three statistical forecasts did better than the five decision-focused variants tested, and on 53 later heavily capped days no variant beat the regression or the average.
Interpretation
In battery arbitrage, clipping forecast prices to a regulatory cap changes the decision even though it lowers forecast error in every block. Clipping looks harmless: realized prices cannot exceed the cap, so cutting forecasts above it only moves them closer to realized values. In a two-period example, clipping lowered mean squared error from 10 to 2 but lowered realized trading profit from 1 to 0, because the flattened spread stopped the battery from trading.
The level a decision-focused loss prefers can be characterized by throughput matching, meaning the energy charged by the helper dispatch problem inside the loss matches that of the perfect-foresight schedule. When the battery starts and ends at the same charge and round-trip efficiency is below one, a common price shift changes each schedule's profit only through charged throughput, so the loss is nearly flat along that direction. At the regression forecast, the pure SPO+ loss-minimizing upward shift rose from 324 to 2,496 rupees per MWh as efficiency went from 0.80 to 0.95, while moving across that whole range lowered daily loss by only 9 to 39 rupees out of about 5,500.
Perspective
The findings come from a small price-taking battery in the Indian day-ahead market, where the cap binds, plus a ranking check in the German day-ahead market; German prices never reach their limits, so that check cannot test the cost of clipping. The theory covers decision problems whose profit is affine in the forecast and whose feasible set does not depend on it and has a convex hull with finitely many extreme points, but the general mechanism is tested empirically only on batteries.
The upward shift depends on the helper problem charging more energy on average than the perfect-foresight schedule, a condition that holds in both markets here but can fail for an overly smooth forecast. The boundedness result with a squared-error term covers only training before projection, and training dynamics have no theory. The safe-clipping certificate is sufficient, not necessary, and gives only an upper bound on the change in regret, with no lower bound that depends only on the share of clipped blocks.