CJ Trowbridge

DS 786 – Week 2

  • Watch Videos
    • Introduction to Forecasting (5:59)
      • Forecast: a statement about the future
        • To be of value, the forecast is about a future event that we don’t have complete certainty for?
        • Given what we know now, how much do we expect to sell next month?
      • Forecast Time Horizons
        • Short-range: immediate future to three months
        • Medium-range: several months to a year
        • Long-range: more than one year
      • Common Aspects of Forecasts
        • Assumes causal system: path => future
          • Do observations of past behavior have relevance in predicting future behavior?
        • Forecasts are never perfect (we just need them to be good enough)
        • Forecasts are more accurate for groups vs individuals
          • (ie sales of the category greek yogurt not just single-serving non-fat strawberry)
        • Forecast accuracy decreases with time
          • Predicting next week’s sales has less uncertainty than predicting sales 50 weeks out.
      • Steps in the forecasting process
        • Step 1 Determine the purpose of the forecast
        • Step 2 Establish time horizon
        • Step 3 Select forecasting technique
        • Step 4 Gather and analyze data in detail
          • State any assumptions
          • Validate data: may need to cleanse of filter for past events
        • Step 5 Calculate forecast
        • Step 6 Analyze/monitor the forecast –
          • Measure accuracy, are results acceptable?
            • No: Return to step 3
            • Yes: Publish forecast
      • Types of forecasts
        • Judgmental
          • Use subjective inputs
            • Consumer surveys, opinions of managers and staff: ie Delphi method
        • Time Series
          • Use historical data and assuming the future will be similar to the past
        • Associative forecasts
          • Use explanatory variables other than time to predict the future.
      • Time series foreacasting
        • Plot your data and look for patterns
        • Trend: Long-term movement in data
          • Linear: steady increase or decrease over time
          • Nonlinear: demand may be exponential, may both increase and decrease over product life cycle
        • Seasonality: short-term regular variations in data
          • What season does walgreens sell the most allergy meds
          • Not just limited to fall/winter/spring/summer variations
            • Which days are busiest for fancy restaurants?
            • Daily cycle at starbucks: when do they sell the most coffee drinks?
        • Irregular variations: caused by unexpected events
          • Infrequent spikes or troughs like toilet paper in 2020
        • Random variations: caused by change
    • Learn how to do a Naive Forecast (4:33)
      • Naive forecasts
        • Stable time-series data
          • eg. sold 50 tires last week, so expect to sell 50 this week
        • Seasonal variations
          • sold 50k heart shaped candy boxes last February so expect to sell 50k this February.
        • Data with trends
          • sold 50 lattes yesterday but 45 the day before so expect to sell 55 today!
      • Naive forecasting: pros and cons
        • Pros
          • You only need one data point
          • Can start soon
          • It’s easy
          • Easy to explain simply
        • Cons
          • Not very accurate
        • Naive forecasts are useful as a benchmark or lower bound for accuracy
    • Working with the Moving Average (4:54)
      • Simple moving average
        • The average of N most recent observations
        • The larger the N, the smoother the forecast, but the greater the lag (ability to respond to “real” changes)
    • How to Work with Weighted Moving Averages and Exponential Smoothing (6:38)
      • Weighted moving average
        • Premise: the most recent observations might have the best predictive value. Yet for simple moving averages, older data points have same importance as most recent
        • Solution: Assign weights to give more importance to more recent observations. (Weights must be between 0 and 1 and all weights must sum to 1 or you get bias)
        • Advantages: avoids over-smoothing and lag time is decreased
        • Weights are often found through trial and error, no one set of weights work best for all situations.
      • Exponential smoothing
        • Exponential smoothing is a type of weighted moving average that considers all time periods that we have data for, but gives them exponentially less weight the further back in time they fall
  • Measuring Forecast Accuracy for Stable Time Series Data (5:23)
    • How to select a forecast
      • for data without trends or seasonality, we could use
        • a simple naive forecast
        • a moving average
        • a weighted moving average
        • an exponential smoother
      • We will get different answers with each technique
        • How do we pick the one most likely to be close to what will happen
      • We calculate past forecast accuracy
    • Measuring forecast accuracy
      • Many different measures exist, such as MSE (mean squared error) and MAPE but in this class we will only calculate mean absolute deviation (MAD)
      • First define an error as the difference between the actual value and predicted value for that time period
      • The take the absolute value of this error
      • Sum these absolute values and take the average

cjtrowbridge

←
→