class: left, top, title-slide .title[ # Predictive Analytics Unit 11: Forecasting Intro ] .author[ ### Ken Arnold
Calvin University ] --- ## Forecasting = Supervised Learning, about the Future ## Time Series - Data observed sequentially over time. - Regular? - We'll discuss *regular intervals*: every second / minute / hour / day / week / month / quarter / year - *Irregular* can also occur (ATM transactions, IoT sensors, storms, outbreaks, sports games, ...) --- ## What Can Be Forecast? Why can phase of moon can be predicted precisely enough to print on calendars, but the stock market can't? - Understanding of contributing factors - physics of large objects, vs complex human social decisions - How much data is available - Lots in both cases - How similar future is to past - moon's orbit is changing very slowly and predictably. - market depends on quickly changing factors: innovations, world events, ... - Whether forecasts will affect what we're trying to predict(!) - Buy/sell recommendations affect stock prices --- ## Variability - Farther in the future => generally more uncertainty (variability) - But models can still be useful. - Models can give *possible future trajectories* ("forecast distribution") - Better model: least "surprised" by actual future - *Prediction interval*: range in which most possible futures occur .pull-left[  ] .pull-right[  ] .floating-source[FPP section 1.7] --- ## How to evaluate forecasts? - Agrees with reality? - How close is actual value to predicted value? - Do the actual value fall within the forecast *range*? - Robust to change? - Understandable to stakeholders? --- ## Time Series Models (vs Explanatory Models) Bike sharing example: - **Explanatory**: rides today = f(temp, day of week, month, is holiday, etc.) + error - **Time series**: rides today = f(rides yesterday, rides last week, etc.) + error - **Dynamic**: rides today = f(rides yesterday, temp, day of week, etc.) + error --- ## Forecasting Process 1. Define problem and stakeholders 2. Gather historical data and expertise 3. EDA 4. Fit and compare models 5. Evaluate model Similar to past units?