How can statistical analysis help you identify significant findings in your data?

Depending on who you are, you will have either of the two types of “views” of an issue - the ant’s eye view or the bird’s eye view. For example, if you live in a village you may know how each of your fellow villagers are doing economically. But you have no idea how people in the hundreds of other villages far from yours are doing. You have an ant’s eye view. On the other hand, if you’re the Governor of the central bank, you know how the country as a whole is doing economically. You have no idea how person X in the village Y is doing. You have a bird’s eye view. Statistical analyses brings together the ant’s and the bird’s eye views. You know the average but you also know the variance, and you can disaggregate as needed.

Statistical analysis is fundamental to almost all analytics I have seen used in the finance industry. The most prominent (and often overused to the point of abuse) is linear regression models. These tend to be the most common go to model for any markets when dealing with two or more variables. In asset management, factor models are themselves rooted in regression. Beyond relational models, another model very often used is optimisation - whether it’s portfolio rebalancing or trade timing, these are fundamental to finance. Over the past decade, I have been using more ML models, but never in isolation, always as a challenger or parallel model to traditional models driven more through SME input into design and selection of variables.

With modern technology and statistical software, you don’t need to go looking for this. Once you know the nature of your data and the output you are interested in all relevant tests are usually already available in the package. For example time series packages in R, Matlab and python are pretty exhaustive. The more crucial element though before building the model, and then reviewing its fit through relevant tests, is understanding the data really well. Not only elements like autocorrelation, but also its market behaviour and interaction, outliers, how to treat missing or erroneous data. Getting this step wrong, the results can be grossly misleading. I often use basic ML (Random Forest) models for exploring data, through randomised inputs

From my experience, models rarely remain static over time, especially since model of the models I have worked on have been on market data. The most important aspect in context of my work has been monitoring stability/variation of factors overtime and what that means in terms of signal generation. A lot of market factors are mean reverting, though reversion cycles may vary, therefore the coefficient shifts is really useful output to monitor to monetise such trades.

80% of models fail to be implemented because they lack easy interpretability in terms of trading behaviours. I’ve found that models that can be simplified to outcomes or actions with tooling that allows quick what if look backs, become much easier to execute and implement across teams. Also unless you are a broker dealer responsible for price pricing, most models can do with lower level precision to get to the same level of success. Over engineering models, often fail because they can’t be easily adapted based on new information, to which SME in markets have intuitive adjustments. In my experience models work most effectively in detecting shifts in patterns, and if that is valuable information to what you do, models can be really powerful