Forecasting the Canadian Dollar Exchange Rate Wissam Saleh & Pablo Navarro
Research Question: What variables effect the Canadian/US exchange rate? Do energy prices have an effect on the Canadian/US exchange rate? 2 views: no significant impact has an impact Before that we discuss the different types of exchange rate regimes and provide a literature review
Exchange rate regimes 1- Fixed exchange rate regimes: Currency is pegged to the US dollar Example: Oil exporting countries such as KSA, Venezuela, and some of other emerging and developing economies) peg their currency to the US dollar Reason: Trade purposes, Less volatility and fluctuations in the economy Central bank intervenes in open market operations to keep the currency pegged to the US dollar
Exchange rate regimes, Cont d 2- Floating exchange rate regimes ( Canada, Germany, USA, France and most developed economies) Currency is not pegged to the US dollar Exchange rate is determined by market forces such as: Supply and demand for the currency, trade Floating exchange rates can response better to shocks in the economy and allow central banks to stabilize other factors such as inflation and unemployment.
Literature Review No impact Ferraro, Rogo and Rossi (2012) - Variables: price of oil, interest rates in Canada and USA only - They use different frequencies ( daily, montly, and quarterly) - lagged oil prices - Non-linear modles - No significant long-run effect
Literature Review, Cont d Has an impact Amano and Norden (1992, 1995) - Variables: price of oil, interest rates in Canada and USA only and they add non oil commodity prices to the model - Use error correction model, a time series model where past deviations from long-run equilibrium influences short-term deviations - they deviations were caused by oil prices and the model had explained most of the variation when the economy was not diversified
How does our model differ? 1- By Incorporating more variables into the model that could affect the exchange rate. Population Growth Unemployment rates Government Expenditure Energy commodity prices Non Energy commodity prices
How does our model differ?, Cont d 2- Using data mining techniques to uncover any relationship that was not reported in previous studies and help trace out the movement in the exchange rate. Using ARMA model ARMA is composed of two models: Autoregressive and Moving Average Used with time series data Predictions depend on past observations/errors The errors are allowed to correlate
Our Variables Month and year. Canadians dollars per US dollar. Real Energy Commodity price index. Real Non Energy Commodity price index. Treasury bills Can: 3 months (Percent). Treasury bills USA: 3 months (Percent). USA Consumer Price Index. Toronto Stock Exchange Composite Index. M1 Gross (Canada most liquid components). M2 Gross (Canada saving bonds, non market mutual funds). Unemployment rate comparable to the US. Expenditure on goods and services. Canada population.
Getting the data Most of the data was obtained in the CanSim website in a CSV format. Data was organized in the csv as one row per month ranging from : March 1986 to September 2015. 1 Output variable (Canadians dollars per US dollar). 13 Input Variables.
Using R for Time series Analysis An R package for forecasting multivariate Time Series was found: dse : Dynamic Systems Estimation. Developed by Paul Gilbert (Bank of Canada). Based on the ARMA model.
Time series data for dse Load the data into a TimeSeries object which has two types of initial data: Input data: variables used to make the prediction. Output data: variables that the model will try to predict Feed the Time series data to the package to generate a model. The package includes a lot of model options to be generated, due to time constraints only a few were tested. The best results were obtained by using the methods to generate VARxLS models. Vector autoregression fitted by least squares regression.
Mar-86 Oct-86 May-87 Dec-87 Jul-88 Feb-89 Sep-89 Apr-90 Nov-90 Jun-91 Jan-92 Aug-92 Mar-93 Oct-93 May-94 Dec-94 Jul-95 Feb-96 Sep-96 Apr-97 Nov-97 Jun-98 Jan-99 Aug-99 Mar-00 Oct-00 May-01 Dec-01 Jul-02 Feb-03 Sep-03 Apr-04 Nov-04 Jun-05 Jan-06 Aug-06 Mar-07 Oct-07 May-08 Dec-08 Jul-09 Feb-10 Sep-10 Apr-11 Nov-11 Jun-12 Jan-13 Aug-13 Mar-14 Oct-14 May-15 The model built using all of our data Exchange Rate 1.6 1.5 1.4 1.3 1.2 1.1 1 0.9 0.8 0.7 Real Data Model
Validation of the model Additional tests were required, because the model was performing well on data it had already used for learning. A new model was built with data from 1986 to 2010. The values from 2011 to 2015 are used to check the accuracy of the model. The input data from 2011 to 2015 is feed to the model. The model forecasts the value of the output data based on the learnings from the input variables.
Forecast performance for the first months 1.06 Exchange Rate 1.04 1.02 1 0.98 0.96 0.94 0.92 0.9 0.88 0.86 Mar-10 Apr-10 May-10 Jun-10 Jul-10 Aug-10 Sep-10 Oct-10 Nov-10 Dec-10 Jan-11 Feb-11 Mar-11 Apr-11 May-11 Jun-11 Jul-11 Aug-11 Sep-11 Oct-11 Forecast Data Real Data
Forecast performance for 5 years 1.4 Exchange Rate 1.3 1.2 1.1 1 0.9 0.8 0.7 Forecast Data Real Data
Forecast performance For the first 10 months the accuracy of the forecasting can be deemed acceptable according to our criteria. After 10 months without real data the accuracy of the model decreases severely.
1.06 Same tests for a model without the energy related variables for 10 months Forecast including energy vs. Forecast with no energy 1.04 1.02 1 0.98 0.96 0.94 0.92 0.9 0.88 0.86 Mar-10 Apr-10 May-10 Jun-10 Jul-10 Aug-10 Sep-10 Oct-10 Nov-10 Dec-10 Jan-11 Feb-11 Mar-11 Apr-11 May-11 Jun-11 Jul-11 Aug-11 Sep-11 Oct-11 Forecast Data Forecast No Energy Real Data
Same tests for a model without the energy related variables for 10 months Forecast including energy vs. Forecast with no energy 1.4 1.3 1.2 1.1 1 0.9 0.8 0.7 Forecast Data No Energy Real Data
Conclusions The variables that were included in this project are correlated to the CAD/USD exchange rate and can be used to do forecasting. The forecasting is more reliable for the 10 following months after the model is generated with the latest data. The Energy prices can be important to do this forecasting but more research is needed. This project was by no means exhaustive.
Future work Do more tests for the model including tests performed by R. Include more input variables in the model to improve accuracy. Obtain a method to generate the output variables with a good degree of accuracy to be able to use this model to do forecasts for future data. Test more R packages that do forecasting multivariate Time Series.
Questions