Now that we've got our cleaned and split dataset, it's time to build a model. Let's walk through exactly what that involves. Here are the five high-level steps that we're going to walk through. First, we need to review our goal or objective, what are we trying to do? What are we trying to predict? Then, before we create the model, you'd want to establish a benchmark for success. Third, you'll select that model type that we want to use. Fourth, as your model gets trained, it'll start outputting performance metrics called loss metrics, which we'll then use to know if our model is as accurate as it can be. Lastly, we'll tune the model and retrain it until we're happy with its performance. So, let's take a specific e-commerce example and establish an objective. Let's say we want to forecast the amount of monthly site visits to our e-commerce website, so then we can better plan when to run certain online promotions. Next, even before you work with any models, let's establish that benchmark for success. Here, we'll specify a certain amount of tolerance for error in our model's prediction, say plus or minus a hundred, or a thousand visits per month. We then use this benchmark to assess whether or not the model is performing as well as we would like it to, can we use the model or not? Then, we're finally ready to choose a model type. Because we're predicting a numeric field, number of site visits, what type of models we try out first? If you're thinking linear regression, that's an absolutely great starting point for forecasting. After the model is trained and evaluated, we'll then need to review those loss metrics from the model, to see if we're comfortable with the model accuracy, against our performance benchmark that we said. MSE mean squared error, or RMSE Root Mean Squared Error, will be our loss metrics for linear regression. Recall that taxicab example that we demoed earlier, we had a plus or minus $12 cab fare discrepancy first off, and we hinted that maybe more feature engineering maybe necessary. Lastly, after your model is trained and evaluated, you can adjust those model knobs called hyperparameters, in additional preprocessing like, add new features or filter out bad data before retraining your model again. Honestly, you'll find that ML as much as an art form here, as it isn't a science to improve your performance. Next up, let's talk BigQuery machine learning specifics, for our forecasting model.
