How to Derive Trading Decisions Systematically
16 Dec 2022
Mario Nakhle, Product Management
Extreme volatilities and the increasing complexities of energy markets are leading a growing number of asset-less and asset-backed power traders to integrate systematic decision making into their routine workflows.
In this article, we’ll take a detailed look at the mechanics of decision-making systems implementation.
How to Create a Decision-Making System?
Trading models are algorithm-based programs which recognize certain patterns in incoming datasets and output signals to trigger buy or sell decisions.
To maximize the accuracy of market predictions, trading models should be tailored to specific use cases, thoroughly backtested, evaluated and continuously monitored once deployed.
1. Spread Selection
The identification of spread opportunities — i.e. price differences between two markets (Day-Ahead auction and Intraday Continuous in Germany, for example) — is the starting point for building a trading model. For this purpose, the following factors need to be assessed:
Potential profits — by evaluating the average absolute value of a spread.
Volatility of spreads — by calculating value-at-risk (i.e. the probability and extent of potential financial losses over a particular time period), expected shortfalls and the likelihood of extreme market events.
Market access requirements — by assessing the technology, asset and liquidity requirements for participating in the corresponding markets.
Market liquidity — by defining volumes of power that can be bought/sold without affecting the market price. This factor is particularly relevant in Intraday markets where the size of bid-ask spreads serves as a strong indicator of liquidity.
Data availability — by ensuring that corresponding data is accessible.
2. Model Design
As a rule of thumb, trading models aren’t supposed to represent all market dynamics — they should rather capture specific market causalities, which drive power prices in a certain direction. Some obvious examples of such causalities are seasonal electricity supply/demand peaks, differences between the expected and actual amounts of solar and wind productions or extraordinarily high fossil fuel prices.
In order to capture market causalities, the following common steps are applied when designing a trading model:
Insights collection. A heterogeneity of fundamental and technical factors which drive the selected spread is analyzed.
Model creation. Based on the collected insights, an algorithm is designed to model the prespecified market causalities and predict the spread direction. At this stage, it’s crucial to avoid overfitting a model against its training data. Commonly, overfitted models fail to derive accurate market predictions from unseen datasets.
Model backtesting. Backtests allow for a model’s viability assessment by testing how well it would have performed on out-of-sample historical data. For the sake of accuracy, it is recommended to only backtest on data with a known publication time. Similarly, regular validations of backtest results based on model predictions are a good practice.
Evaluation of model performance. Typical model performance metrics to look at are profit & loss per MWh, value-at-risk, a risk-adjusted return measure called annualized Sharpe ratio and drawdowns to depict peak-to-trough difference of a particular trading account/position over a specific time interval.
Signals execution. The last stage of model creation involves determining technical requirements for the execution of signals. In power markets, particular attention has to be paid to Intraday execution since an expensive execution of orders in Intraday markets can render an otherwise good signal unprofitable.
3. Position Sizing
Before running a trading model, it’s crucial to define one’s maximum acceptable loss and apply common risk management techniques — i.e. conduct the above-mentioned value-at-risk, drawdowns and standard deviation analyses. This helps determine trading volumes and set realistic expectations.
4. Model Monitoring
After a trading model is up and running, its performance shall be monitored continuously:
Evaluate live and historical performances to detect overfitting and adjust a trading model to market dynamics when necessary.
Analyze distribution of data inputs in a running model compared to a training data distribution to examine how quickly a given model adapts to sudden market changes.
Monitor the explainability of the trading model to understand the logic behind decisions. A lack of transparency can lead to building a trading strategy based on false assumptions which often results in trading performance instabilities.
Monitor compliance with risk metrics set at the “Position Sizing” step to make sure a model operates within its defined limits. If those limits are exceeded, intervention is required.
Automation of decision making aims at excluding emotions from the trading process by replacing human judgement with decision-making systems.
While the development of such systems requires considerable technical effort, it allows energy market participants to streamline their decision making, mitigate human errors and increase trading performance.