Forecasting the LFS. Fida Hussain Nick Misoulis Nigel Stuttard

Transcription

1 Forecasting the LFS Fida Hussain Nick Misoulis Nigel Stuttard

2 J K Galbraith The only function of economic forecasting is to make astrology look respectable

3 Overview Current estimates, rolling three-month estimate, produced 6 weeks after end of quarter. Need more timely estimates of key labour market indicators, particularly unemployment to meet PEEI requirements. Can forecasting help? Forecasting monthly data Forecasting weekly data

4 Structure of LFS Continuous household survey, covering 60,000 households per quarter. Sample consists of 5 waves ; each wave is interviewed over 5 successive quarters. Sample allocated to 164 interviewing areas. Each area split into 13 stints ; each stint area randomly allocated to one of 13 weeks of each quarter.

5 Constructing a monthly series Weighting replication of system used for weighting full LFS Seasonal adjusting common ARIMA model (210)(011), 3x5 seasonal moving average, 23 point Henderson moving average Constraining to ensure additivity of final series Benchmarking To ensure that average of 3 single months is same as published moving 3-month average

6 How single month series are used Employment Rate Dec 03 Dec 05 Chart 1: UK Working Age Employment Rates (Seasonally Adjusted) Rate (%) Single month Published LFS

7 How single month series are used Employment Rate Feb 04-Feb 06 Chart 1: UK Working Age Employment Rates (Seasonally Adjusted) Feb-04 Mar-04 May-04 Jul-04 Sep-04 Nov-04 Jan-05 Mar-05 May-05 Jul-05 Sep-05 Rate (%) Nov-05 Jan-06 Feb-06 Single month Published LFS

8 Models chosen for forecasting single month series All Under 25 Over 25 Male Employment 210+c c Male Unemployment 210+c (210)(001) 210+c Male Inactivity 210+c c Female Employment 211+c c Female Unemployment 210+c (011)(002)+c 213+c Female Inactivity (210)(001) 212 (210)(001)

9 Performance of single month forecasts Mean Error (ME %) Single 1 month ahead Forecast Root Mean Square Error (RMSE %) Thiel s U Statistic (U) Single 2 month ahead Forecast Mean Root Thiel s U Error Mean Statistic (U) (ME %) Square Error (RMSE %) Male Employment 0.146* * Male Employment < Male Employment > Male Unemployment 2.256* * Male Unemployment < Male Unemployment > * * Male Inactivity Male Inactivity < * Male Inactivity > Male Activity * Female Employment Female Employment < * Female Employment > Female Unemployment Female Unemployment < * * Female Unemployment > Female Inactivity Female Inactivity < * * Female Inactivity > Female Activity

10 Single month forecasts - male Male Employment Levels Male Employment Rates 15, % 15,550 15,500 15,450 15,400 15, % 66.9% 66.8% 66.7% 66.6% 66.5% 15, % IFR 1 Month IFR 1 Month Male Unemployment Levels Male Unemployment Rates % 5.5% 5.4% 5.3% 5.2% 5.1% 5.0% 4.9% 4.8% 4.7% 4.6% IFR 1 Month IFR 1-month-ahead forecast

11 Single month forecasts -female Female Employment Levels Female Employment Rates 13,400 13,350 13,300 13,250 13,200 13,150 13,100 13,050 13, % 53.8% 53.6% 53.4% 53.2% 53.0% 52.8% IFR 1 Month IFR 1 Month Female Unemployment Levels Female Unemployment Rates % 4.6% 4.5% 4.4% 4.3% 4.2% 4.1% 4.0% 3.9% 3.8% IFR 1 Month IFR 1 Month

12 Forecasting weekly data Utilise two / three weeks more data. Production system not yet able to deliver weekly data in time. Weighting of weekly data more problematic.

13 Modelling weekly data 2πkx( t) 2πkx( t) [ y ARIMA( p, d, q)( P, D, Q) p ' t zt g ( ak cos + bk sin )] = k = 1 p p 13 where y t is the weekly series, z t is a vector of regression variables other than the trigonometric components, g is a vector of unknown coefficients, and the sum is the trigonometric components used to describe the seasonal pattern. We need to estimate the parameters g,a and b. Also fit the appropriate ARIMA model.

14 Models fitted to the weekly time series Series Selected model Probability of Q at lag ARIMA Trigonometric frequencies Female Unemployment (011)(100) 1,2,3,4, Male Unemployment (011)(100) 2,3,4, Female Employment (311)(100) 2,3,5, Male Employment (011)(100) 1,2,3, Female Inactivity (300)(100)* 1,2,3, Male Inactivity (011)(100) 1,2, * The model also includes a constant, while it does not include the terms AR(1) and AR(2).

15 Comparing empirical performance of different forecasts Series Modelling Forecasting Forecasting weekly data quarter monthly data Weekly Monthly (benchmark) Monthly modelling modelling modelling Mean % RMSE % Mean % RMSE % Mean % RMSE % Mean % RMSE % F male Unemployment Male Unemployment Female Employment Male Employment Female Inactivity Male Inactivity

16 Conclusions Early estimates based on forecasting single month series better than those based on forecasting quarterly series. Unemployment estimates based on weekly data better than those based on monthly data. Employment and inactivity estimates not noticeably better when based on weekly data.