Business Forecasting Topic 4 Flashcards
Global models of time series
assume same trend or seasonal pattern applies to entire time series
local models of time series
assume different trends or seasonal patterns can apply at different points in time
exponential smoothing models
allow forecasts to take into account any changing patterns by updating estimates of patterns as new observation is available
- older/later data = less weight
Simple (or single) exponential smoothing
- where series has flat trend
- subject to changes to new levels
- estimate current underlying level of time series (level when noise filtered out)
- estimate of current lvl = forecast for next period
- smooth out random fluctuations in series (not eliminate randomness)
- only works if not a trended series otherwise all same values
SES forecasts formulae
ON SHEET!!
forecast = weighted average of latest observation and current forecast
- larger α = greater weight attached to most recent observation
more than 1 period ahead = one period ahead forecasts
initial forecast for SES
required to start process
- average of past observations
- educated guess
smoothing constant
α
0 = smooths thoroughly
1 = naive forecast
- value chosen to balance stability (not overact to freak figures) and sensitivity (respond to new conditions)
- α value with lowest MSE = produce future forecasts
- MSE over the fitting period, find optimal α to minimise MSE
high α value
α = 1 = naive forecasts = shift 1 period
behave exactly like data
respond to sensitivity
low α value
dont catch up data
Exponentially weighted moving averages (EWMA)
each forecast = weighted average of all previous observations
most recent = highest weight
oldest = lower weight
forecasting more than 1 period ahead
SES assume no upward or downward trend in time series = extrapolate flat line in future
- forecasts of more than 1 period ahead are identical to the one period ahead forecast
advantages of SES
- simple
- easily automated
- low data storage req
- more accurate than many complicated methods
- large number forecast on regular basis
tracking signals
- detect when forecasts = inaccurate
- α value needs to be adjusted
error high = tracking signal high = showing a widening gap between data and forecasts
cumulative error
- if growing -> ideally values close to zero if unbiased forecasts
- alerts us when things are wrong
issues with using cumulative error as method of alerting of issue
- how large/small (small = -ve) does error have to be before alert
- one large/unrepresentative error keeps it high/low -> even if subsequent forecasts are unbiased