💰 quantitative finance

High-frequency intraday trading for battery storages

This paper presents a computationally efficient, high-frequency dynamic programming approach for optimizing grid-scale battery storage in continuous intraday electricity markets, demonstrating that millisecond-level re-optimization significantly boosts revenue compared to slower strategies and enables effective parametric training.

Original authors: David Schaurecker, David Wozabal, Nils Löhndorf, Thorsten Staake

Published 2026-08-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: David Schaurecker, David Wozabal, Nils Löhndorf, Thorsten Staake

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the electricity grid as a giant, humming nervous system for a modern city. Just like your body needs to keep its temperature steady, the grid needs to keep the flow of power perfectly balanced: if you use too much electricity, the lights flicker; if you produce too much, the system gets overloaded. To fix this, we have massive batteries that act like giant sponges, soaking up extra energy when the sun is shining or the wind is blowing, and squeezing it out when the demand spikes. But here's the tricky part: these batteries aren't just sitting there waiting to be told what to do. They are traders. They need to buy electricity when it's cheap and sell it when it's expensive, often doing this thousands of times a day to make a profit.

The market where this happens is called the "intraday market." Think of it as a super-fast, high-stakes video game where the price of electricity changes every single millisecond based on a massive list of buy and sell orders, known as a "limit order book." It's like a digital auction house where the rules change faster than you can blink. The big question for scientists and engineers is: How do you program a battery to play this game perfectly? If the battery is too slow, it misses the best deals. If it's too fast but makes the wrong moves, it wastes money or wears out the battery too quickly. This paper dives into the heart of that problem, asking if we can build a trading brain for batteries that is fast enough to catch every tiny price change without crashing the computer.

The authors of this paper, a team of researchers from Switzerland, the Netherlands, Luxembourg, and Germany, decided to tackle this by creating a new, super-fast way to tell a battery how to trade. They took a standard trading strategy called "rolling intrinsic," which is like a smart shopper who constantly checks the price tags in a store and buys or sells based on what's available right now. However, the old way of doing this math was like trying to solve a giant, complex puzzle with a calculator that takes hours to crunch the numbers. By the time the calculator finished, the prices in the store had already changed, making the answer useless.

To fix this, the researchers invented a new method using something called "dynamic programming." Imagine you are navigating a maze. The old method tried to calculate every single possible path through the maze at once, which took forever. The new method breaks the maze down into tiny, manageable steps, solving each step instantly and moving forward. This allowed them to run their battery trading simulation at a speed that matches the real market: they could make a decision in a few milliseconds, reacting to every single update in the order book.

When they tested this new, lightning-fast strategy on a full year of real market data from Germany, the results were eye-opening. They found that speed is everything. Their high-frequency strategy, which re-thinks its plan every time the market updates (which happens thousands of times a second), earned 58% more money than a strategy that only re-thinks its plan once an hour. Even compared to a strategy that updates once a minute, their fast version made 14% more profit. This proves that in the world of electricity trading, being fast isn't just a nice-to-have; it's the difference between making a fortune and leaving money on the table.

The paper also showed that their fast method didn't just guess; it was incredibly accurate. They compared their new "quick and dirty" math against the "slow and perfect" math (which takes hours to solve) and found that the fast version made almost the exact same amount of money, just much, much faster. In fact, their algorithm was so fast that they could run a full year of simulations in about 86 minutes, solving the trading puzzle roughly 24 million times.

Furthermore, because their method was so fast, they were able to tweak the strategy's settings to make it even better. By adding a simple "penalty" to discourage trading when the market was too messy or slow, they managed to squeeze out an extra 8.4% profit on top of what they were already making. This suggests that with the right tools, we can teach batteries to be not just fast, but also smart enough to know when not to trade.

However, the authors are careful to point out that this is a simulation. They used real data from 2021, but they didn't actually put a battery on the market. They also simplified some of the battery's physical wear and tear to keep the math fast, and they didn't account for how their own trading might have changed the behavior of other traders in the real world. Despite these limitations, the study strongly suggests that the future of battery trading lies in algorithms that can think and act at the speed of the market itself. If we want to get the most out of our renewable energy and keep the grid stable, we need to stop waiting and start trading at the speed of light.

Technical Summary: High-Frequency Intraday Trading for Battery Storages

Problem Statement
Grid-scale battery energy storage systems (BESS) require sophisticated trading strategies to maximize revenue in continuous intraday electricity markets. While the "rolling intrinsic" (RI) strategy is a standard benchmark for storage operation, its application to continuous intraday markets is hindered by computational constraints. Existing literature often relies on stylized market models, aggregated transaction data, or discrete time steps (e.g., hourly or 15-minute intervals), which fail to capture the millisecond-level dynamics, liquidity constraints, and microstructure of continuous order books. Furthermore, exact optimization methods, such as Mixed-Integer Linear Programming (MILP), are too computationally expensive to re-optimize at the frequency required to react to every relevant order book update. This gap prevents realistic backtesting and the exploitation of short-lived arbitrage opportunities that depend on trading speed.

Methodology
The paper proposes a computationally efficient, high-frequency implementation of the rolling intrinsic strategy tailored for continuous intraday markets. The methodology involves three core components:

Full Limit Order Book (LOB) Representation: Unlike approaches using aggregated price curves, the model explicitly incorporates the full limit order book, including all active buy and sell offers, order types, and market rules. It integrates a linear approximation of battery degradation costs and trading fees directly into the objective function to prevent economically unviable marginal trades.
Dynamic Programming (DP) Reformulation: The authors reformulate the standard MILP intrinsic optimization problem into a multi-stage dynamic programming problem. While the physical decisions for all delivery hours are made simultaneously in reality, the DP approach maps each delivery hour to a distinct "stage," using the battery's state of charge (SoC) as the connecting state variable.
- The value function is approximated by discretizing the SoC on a finite grid and using linear interpolation.
- This reformulation trades a marginal loss in theoretical precision for a computational speedup of two to three orders of magnitude compared to exact MILP solvers.
- The algorithm executes backward induction to determine optimal value functions and a forward pass to derive actionable trading decisions.
High-Frequency Execution Logic: The strategy is designed to re-optimize at every relevant order book update (millisecond resolution). To manage execution risk and latency, the system submits "fill-or-kill" orders. If an order is missed due to market shifts during the solve and technical delay, the rolling nature of the policy automatically corrects the resulting non-physical positions in the subsequent optimization step.

Key Contributions
The paper makes five primary contributions to the field of energy trading and operations research:

Adaptation of Rolling Intrinsic to Continuous Markets: The authors adapt the RI strategy to explicitly model every single order in the limit order book, formulating the optimization as a MILP that accounts for order placement costs, linear degradation, and detailed market rules.
DP Approximation for Speed: They introduce a DP approximation that achieves a speedup of 100–1000 times over standard MILP solutions. This enables the first high-frequency algorithmic trading strategy for storage assets capable of reacting to every relevant order book update in real-time.
Comprehensive Backtesting: A detailed backtest over the full year of 2021 using German order book data (240.4 million order submissions) demonstrates that the DP formulation does not reduce trading profits compared to the exact MILP solution while enabling realistic, rapid simulation.
Parametric Optimization: Leveraging the algorithm's speed, the authors train a parametric extension of the RI (introducing a penalty for wide bid-ask spreads) to optimize trading behavior, achieving an additional 8.4% revenue increase out-of-sample.
Open-Source Tool: The authors release a Python package (BitePy) to facilitate the simulation of high-frequency rolling intrinsic trading scenarios for researchers and industry practitioners.

Results
Numerical experiments on 2021 German intraday data yield the following findings:

Revenue Impact of Speed: Trading speed is critical. The high-frequency policy (re-optimizing at every relevant LOB update) earns 58% more revenue than a strategy re-optimizing once per hour and 14% more than a strategy re-optimizing once per minute.
DP vs. MILP Performance: The DP approach with a coarse grid ( $m=11$ ) achieves profits comparable to the exact MILP solution (within statistical significance) but reduces simulation time from hours to minutes (e.g., 86 minutes for a full year vs. significantly longer for MILP).
Execution Friction: Approximately 7% of orders are missed due to latency (solve time + technical delay). The rolling policy corrects these infeasibilities rapidly (median 0.3 seconds), resulting in a total profit reduction of roughly 6–7% due to these frictions, yet the high-frequency strategy still significantly outperforms lower-frequency benchmarks.
Parameter Sensitivity: Finer storage discretization ( $m$ ) yields marginally higher profits. Higher degradation costs and lower efficiencies reduce total rewards. Interestingly, smaller storage capacities yield higher revenue per MWh due to the ability to avoid sweeping deep into the illiquid order book.
Parametric Tuning: Introducing a penalty parameter for illiquid products (wide bid-ask spreads) and optimizing it via a sliding window approach increased yearly revenue by 8.4% compared to the standard unpenalized RI.

Significance and Claims
The paper claims that its work establishes a critical foundational step for understanding and modeling the continuous intraday market. By isolating this market and providing a highly accurate, fast-executing benchmark, the authors argue that previous studies relying on artificial time discretization have yielded suboptimal strategies and biased revenue estimates.

The central finding is that the continuous clearing mechanism of intraday markets structurally rewards high-frequency execution. The authors demonstrate that strategies failing to react to tick-level order book updates leave significant value on the table. The proposed method proves that it is possible to model the full complexity of the market (microstructure, degradation, rules) and execute decisions at millisecond resolution, a capability previously deemed computationally intractable. This enables realistic backtesting and the development of more sophisticated, data-driven trading strategies that were previously impossible to train or evaluate. The authors position this as a necessary precursor to expanding such algorithms into broader multi-market environments (e.g., combining intraday with day-ahead or reserve markets) without sacrificing intraday realism.

Technical Summary: High-Frequency Intraday Trading for Battery Storages

More like this