Predictive Analytics
Predictive Analytics is the sub-field of Supervised Learning where users attempt to model data elements and predict future outcomes through evaluation of probability estimates. Rooted deeply in mathematics specifically statistics, predictive analytics shares many components with Unsupervised Learning, with the prescribed difference to the measurement of a desired predictive outcome.
Predictive Analytics is the development of probability models based on variables, including historical data, related to possible events (purchases, changes in price, etc.). When it receives other pieces of information, the model triggers a reaction by the organization. The triggering factor may be an event, such as a customer adding a product to an on-line shopping basket, or it may be data in a data stream, such as a news feed or utility sensor data, or an increased volume of service requests. The triggering factor may be an external event. News being reported about a company is a big predictor of a change in stock price. Predicting stock movement should include monitoring news and determining if news about a company is likely to be good or bad for the stock price.
Frequently, the triggering factor is the accumulation of a large volume of realtime data, such as an extremely high number of trades or requests for service or volatility of the environment. Monitoring a data event stream includes incrementally building on the populated models until a threshold is reached as defined in the model.
The amount of time that a predictive model provides between the prediction and event predicted is frequently very small (seconds or less than a second). Investment in very low latency technology solutions, such as in-memory databases, high-speed networks, and even physically proximity to the source of the data, optimizes an organization’s ability to react to the prediction.
The simplest form of predictive model is the forecast. Many techniques exist for trending or forecasting based on regression analysis and benefit from smoothing. The simplest way to smooth data is through a moving average, or even a weighted moving average. More advanced techniques can be useful, like the exponential moving average, which introduces a smoothing factor to be applied. Minimizing the error residual from the least squares can be a starting point, but several runs are necessary to determine and optimize the smoothing factor. Double and triple exponential smoothing models exist to address trend and seasonality components.
Prescriptive Analytics
Prescriptive analytics take predictive analytics a step farther to define actions that will affect outcomes, rather than just predicting the outcomes from actions that have occurred. Prescriptive analytics anticipates what will happen, when it will happen, and implies why it will happen. Because prescriptive analytics can show the implications of various decisions, it can suggest how to take advantage of an opportunity or avoid a risk. Prescriptive analytics can continually take in new data to re-predict and re-prescribe. This process can improve prediction accuracy and result in better prescriptions.
