In the previous section, we discussed the factors that causes profitable trading strategies to die, which include because the predictive power of trading signals died out over time, either due to lack of parameter optimizations, discovery by other market participants, violations of underlying assumptions, or seasonal trends. Before we explore optimizing trading signals and what that pipeline looks like, one component that is an important part of any quantitative research platform is called the trading signals dictionary/database. This component is a large database containing statistics of different trading signals and different trading signal parameter sets over years of data.

The statistics that this database contains are primarily ones to capture the predictive abilities of these signals over their prediction horizon. Some simple examples of such metrics can be the correlation of the trading signal value with the price movements in the trading instrument which this trading signal is meant for. Other statistics can be variance in the predictive power over days, that is, how consistent this trading signal is over a set amount of days to check whether it varies wildly over time.

In this database, there can be one entry per day or multiple entries per day for different time periods for every <signal, signal input instruments, signal parameters> tuple. As you can imagine, this database can grow to be very large. Sophisticated algorithmic trading participants often have database results going back several years for thousands of trading signal variants as well as complex systems to compute and add entries to this database with every additional day of market data recorded. The main advantage of having such a database is that, as market conditions change, it is very easy to query this database to understand and analyze which trading signal, signal input, and signal parameter sets do better than others in different market conditions. This helps us to analyze why certain signals might not be performing well in current market conditions, see which ones would have done better, and build new and diverse trading strategies based on those observations.

In a lot of ways, having access to a comprehensive trading signal dictionary/database allows us to quickly detect changing market conditions/participants by comparing the trading signal performance individually across training and testing history to see whether it is deviating from historical expectations. It also helps us to adapt to changing market conditions/participants by letting us quickly query the database for historical signal performance to see what other signals would have helped or worked better. It also answers the question of whether the same trading signal with same trading instrument input, but with different trading signal parameters, would have done better than the current parameter set being used in live trading.

Investing in setting up a research-platform component that can compute results across different trading signals, signal instrument input, signal parameters, signal prediction horizon, time periods over years of tick data, and then storing it in an organized manner can help you to understand and handle a lot of the factors that cause trading-signal-profit decay in algorithmic trading strategies deployed to live markets and facing changing market conditions.