Problem Description:
Unemployment benefits are a huge cost to governments and are highly dependent on
projected unemployment rate for the country. Unemployment benefits are social welfare
payments to unemployed individuals. The definition of unemployed individual varies from
among different governments. Each year Governments allocate a certain percentage of
their financial outlay to meet these requirements.
Forecasting Objective:
Forecasting unemployment rates is critical for the budgetary allocations. We forecast
monthly unemployment rates for 5 different countries that have income support as part of
their social welfare scheme. We choose monthly unemployment rates as generally these
payments are monthly.
Client Information:
Our Clients are Ministries of Finance of European and UK governments who have to budget
for unemployment benefits as part of their social welfare schemes.
Data Description:
- Source: Federal reserve of Economic Data (https://fred.stlouisfed.org/)
- Key Characteristics: Trend, Level and noise observed for all the data.
- Countries to be analyzed: Austria, UK, Ireland, Germany
Forecast period:
The forecast period is chosen as 3 months + 12 months as the client expects us to forecast
the results 3 months prior to the start of the Fiscal year.
- Forecasting Horizon -> 15 months
- Seasonality -> 12 months
Final Forecasting Method:
Most of our time series had trend and Seasonality. Hence, we used Holt-Winters, MLR and
MLR + ARIMA models for forecasting the time series.
Recommendations:
- For most countries we tried many methods of which MLR with lag 1 was highly accurate with MAPE of under 2%
- However, MLR is difficult and more costly to implement. So, one can use Holt-Winters to forecast as well.
- We recommend sensitising the forecasted values using the confidence interval. This should help the government adjust for buffer allowances
Limitations:
- No external economic Indicators were used.
- The model is highly dependent on the frequency of collection of unemployment data.
- The MLR model uses Lag-1 as an input variable
- Lag-1 data will not be available ahead of time
- Allows us to forecast only one month at a time
Alternatives:
- Use lag-13 if Naïve forecast has reasonable MAPE