University of Illinois at Chicago
College of Business Administration

MBA 503:   Statistics
Instructor:   Prof. Stanley L. Sclove
Textbook:   Levine, Berenson, Stephan(LBS), updated ed. (1998)

Notes on Time Series Analysis to Accompany LBS Ch. 13
Summary

These notes Copyright © 1998 Stanley Louis Sclove
Only Sections 13.1, 2 and 6 are required.
Program 6--"Time Series" from the series Against All Odds: Inside Statistics can be viewed as part of an introduction to time series analysis.)

A time series is a set of numerical data obtained at regular periods of time.

Forecasting is the use of time series analysis and other methods for prediction. Business Forecasting is important; many economic and business datasets are time series. Areas of application include marketing, banking, finance and investment. In Operations Management, time series analysis plays an important role in forecasting demand, planning production and controlling inventory.

Autoregression

The correlation between observations k time intervals apart is numerically measured by the lag k autocorrelation coefficient. This is just like a correlation between an X and a Y, where Y is Yt and X is Yt-k.

If one or more of the autocorrelation coefficients is high, an "autoregressive" model is fitted. A first-order autoregressive prediction equation is

^Yt+1   =   b0 + b1 Yt .

A second-order model would include also Yt-1; e.g., tomorrow's result would be predicted using not only today's but also yesterday's.

The Durbin-Watson statistic is approximately 2(1-r) where r is the 1st-order autoregression coefficient of the residuals. Since 0 < r < 1, DW is between 0 and 4. A value DW = 0 corresponds to r=1; DW = 4, to r = -1. DW = 2, to r = 0 . The residuals are supposed to be uncorrelated, so r = 0 (DW = 2) is ideal. If DW is too far from 2 (roughly speaking, as a rule of thumb, less than 1.5 or more than 2.5, although these limits really depend upon n ), a different model should be tried: Probably some important variable has been omitted.

Statistical computer packages include the ARIMA (AutoRegressive Integrated Moving Average) command, which will fit time series models easily. Time series which are trending up or down should be differenced. Then, which model to use is indicated by the pattern of autocorrelations and the corresponding partial autocorrelations (autocorrelations with the effect of intervening lags removed).

Exponential Smoothing is included in many business software packages. The exponential smoothing forecasting scheme is
^Yt   =   W Yt-1 + (1-W) ^ Yt-1 .
Each new forecast requires only the current value of the series and its forecast.

Seasonal models regress, for example, this quarter's Y on that for the corresponding quarter in preceding years. Two-way tables are useful for presenting seasonal data; e.g., with the rows being months and the columns being years.

Transfer function models relate a time series of Y to that of one or more X's. For example, since steel is made from iron and coal, it makes sense to try to predict steel prices from iron and coal prices. Let
St = this quarter's price of steel,
It = this quarter's price of iron, and
Ct = this quarter's price of coal.
One might consider a transfer-function model

^St =   a0 + a1It-1 + a2Ct-1

Grainger Causality

However, one would not really want to say that I and C determined or "caused" the price of steel (S) unless one took account of the possibility of predicting St from St-1. Hence one would consider a model

^ St = b0 + b1St-1 + b2It-1 + b3Ct-1.
Along with this model one might consider the simultaneous equations
^ It = c0 + c1It-1
and
^ Ct = d0 + d1Ct-1.
Intervention models describe the effects on a time series of specific events (passing of a seat-belt law, passing of a law regarding gasoline composition, an ad campaign, an OPEC meeting) can be modeled using "intervention analysis." A model such as
Yt = Bxt + zt,

where xt has the value 0 before the intervention and 1 afterward, and zt is an appropriate time series, can be used. Or, the response to intervention may be modeled by a more complicated response function than this simple shift.

FURTHER READING

Box, George E.P., Hunter, Wm.G. and Hunter, J. Stuart (1978). Statistics for Experimenters: An Introduction to Design, Data Analysis and Model Building. John Wiley & Sons, New York. Chapter 18 is on time-series analysis.
latest revision 11-Oct-1998