Problems of retraining

An example of retraining or re-optimization when setting up trading systems

Algo traders and developers face the phenomenon of overoptimization when selecting the parameters of automated trading systems. Overoptimization is a process in which models or strategies become overly adapted to historical data and lose their effectiveness on new data.

Overoptimization of automated trading strategies can occur for several reasons:

Striving for high results based on historical data: Traders may adjust their models or strategies too carefully to match past data perfectly. This may lead to the model working well on historical data, but adapting poorly to new market conditions.
Using complex models: Sometimes trading system developers use very complex mathematical models or algorithms that may be too sensitive to small changes in the data. Such models often turn out to be ineffective in real trading conditions.
Lack of data or its quality: if the data on which the model is trained is limited or of poor quality, this may force parameter selection algorithms to resort to overoptimization in order to achieve acceptable results.
Market instability: Financial markets are very dynamic and are influenced by many factors. Strategies that have worked well in the past may become ineffective due to changes in market conditions.

It is important to understand that overoptimization can lead to a deterioration in the effectiveness of strategies in practice and an increase in risks.

Ways to solve the problem of overoptimization

The irony of the re-optimization task is that there will be a lot of results in the optimization results that will show excellent results in the future. This is a fact, because if after 5 years we try to optimize our system throughout the site, we will certainly find the parameters that form an even, upward curve of our account balance. Of course, this applies only to systems with sufficient adaptability. It turns out that we already have the necessary (able to earn on new data) parameters for our trading system – we just need to select them)

Expanding the data sample provides a more representative picture of the system's behavior in various market conditions, which reduces the risk of overfitting. You can use additional characters or just a deeper history of the character.
Parallel testing of several sets of parameters with small deviations helps to identify optimal values and determine the zone of system resistance to changes in input parameters. That is, use several sets of parameters at once and test several systems simultaneously, shifting the parameters in both directions. For example, if we are trying to select a moving average period and test parameter 100, then two more identical systems with parameters 95 and 105 should participate in testing at the same time.
The use of special optimization criteria focused on operational stability makes it possible to select the most reliable system configurations capable of showing stable results in real trading conditions.

Thus, although it is impossible to completely eliminate overoptimization, the application of an integrated approach to parameter selection and evaluation of results makes it possible to create a trading system capable of adapting to changing market conditions and demonstrating stable results in the long term.