August 2021 Summaries

As someone who has built reverse ETL jobs before tooling for this type of work was largely available, implementing bespoke reverse ETL implementation can lead to ongoing maintenance time snowballing as APIs change, new fields are added, or data grows. Adopting a pre-built reverse ETL tool like Census can automate an otherwise grueling task and save time, sanity, and engineering effort. A dedicated reverse ETL tool empowers the analytics team to focus on business needs rather than technical specs, enabling organizations to send branded and highly personalized communications thanks to the data sent from their warehouse. It's essential to loop in stakeholders early and often, understanding the strategic importance of the data your stakeholders need and finding the easiest and most scalable way to send it from A to B. Collaboration between analytics and business teams is key to defining specs of the reverse ETL process, and communication channels should be kept open post-implementation to ensure a smooth experience for all parties involved.

Fraud prediction models are becoming increasingly important as the world becomes more digitized, with record-highs of fraudulent activity reaching $42 billion in 2020. There are four types of fraud models: profile-specific models that identify fraudulent users, transaction-specific models that identify fraudulent transactions, rules-based models that use hard-coded rules to detect fraudulent activity, and machine learning models that learn from features to identify signals of fraudulent behavior. When choosing features for a model, it's essential to include as many signals indicating fraudulent activity as possible. Evaluating a fraud model is different from normal machine learning models due to the significant imbalance between fraudulent and non-fraudulent profiles/transactions, making accuracy an unsuitable metric. Instead, precision and recall metrics are used to assess the model's performance, with precision being suitable for scenarios where the cost of classifying a non-fraudulent transaction as fraudulent is high, and recall being critical when identifying every single fraudulent transaction.