Forecasting Flashcards
What are the two main types of forecasting?
Extrapolation methods- involve the analysis and manipulation of historical data to find historical patterns.
Judgemental forecasting- uses human judgment where data is scarce or volatile.
What is a major assumption of extrapolation forecasting?
Underlying patterns/ relationships found in the data will not change during the forecasting period.
What is the principle of parsimony?
Build the simplest model possible that still accomplishes the task.
What are the forecasting steps?
1) Define the problem
2) Collect information
3) Preliminary analysis
4) Choose and implement forecast methods
5) Use and evaluate the forecasting model
What is an important driver of the values taken by the measure of interest?
Time
What is extrapolation forecasting?
Extrapolating underlying historic patterns and/or existing relationships in data into the future in order to predict future values.
What is time series data?
Data is collected and recorded over successive increments of time. Observations must be displayed in time-order and they are generally not independent.
Important characteristics of time series data
1) Trends
2) Cycles
3) Increasing variance
4) Seasonality
5) Outliers
What might we not want to capture in the model?
Randomness. We do not want to fit our model to this randomness as it would bias our future forecast.
General quantitative forecasting model fitting procedure.
1) Cut off a portion of the data (preferably the amount you want to forecast).
2) Create your model using the initialization data (the data not cut off).
3) Check how good the model is using the cut-off data points.
4) Once you’re happy with the model refit it to the whole data set.
What is white noise?
A stationary time series in which the obersevations are independent and the mean is zero.
What does seasonality look like?
Repeated patterns.
What is ex-post analysis?
Using the fitted forecast model to forecast into the most current part of the observed time series data.
What issues may occur when fitting and using a simple regression model?
1) The linear line fitting to the trend will not fit to any cycles in the data set.
2)You are forecasting outside the range of the observed data, so you have to assume that the historic trend will continue into the future forecast period.
3)If you do not fit exactly to the important characteristics in the time series you may get auto-correlated residuals.
4)The trend may not always be linear in nature, and a simple regression (one explanatory variable) will not fit well to non-linear trends.
What is autocorrelation?
When autocorrelation is present, successive data observations are correlated with each other. Can be seen when residuals are not randomly scattered.
Negatives of autocorrelation
Occurs when regression is used with time series.
1) Affects the accuracy of variance estimates.
2) Affects the accuracy of t-tests and F-tests.
What is positive autocorrelation?
Large values tend to be followed by other large values and small with small.
What is negative autocorrelation?
Large values are followed by small values and visa versa.
How to spot autocorrelation?
Using the Durbin- Watson statistic.
How to interpret DW statistics?
They range from 0 to 4 with a middle value of 2.
A small DW value indicates positive autocorrelation.
A large DW value indicates negative autocorrelation.
DW value near 2 indicates no autocorrelation.
What does lag 1/ first order autocorrelation mean?
The DW stat only looks at neighboring points and the correlation between those.
What is seasonal autocorrelation?
Where there is a correlation between residuals more than one time period apart. E.g. repeating every few months.
What are causal forecasting models?
When we add other variables that we believe have a relationship with the response variable into the regression model.
What is a prediction interval?
This shows the likely values that the forecasted measure might actually take.
What is an interval forecast?
A PI along with a point forecast it is placed around.
When is a PI valid?
Errors must be assumed to be independent and normally distributed with zero mean.
