Impressive Financial Forecasting with SVR

A couple of researchers (Li, Hu, and Hirosawa) have displayed pretty interesting findings forecasting the USD/JPY exchange rate via a Support Vector Regression (SVR) network. The file explains the details. Let's break down what they did and accuracy results from the empirical test.

Basic idea of SVR: Nonlinear mapping of data into a high dimensional feature space with a kernel function, then do a linear regression in the transformed space. It all results in a nonlinear regression in low dimensional space.

Inputs of their study:

f(t): Historical USD/JPY exchange rate

F(t): Historical USD/EUR exchange rate

N(t): Historical NIKKEI 225 stock index value

O(t): Historical oil price in USD

Predicted Value (Output): USD/JPY exchange rate at the next day (t+1), next week (t+5), and 2 weeks from now (t+10)

Prediction evaluation tools:

Sum of Square Error (SSE)

Mean Absolute Error (MAE)

SSE and MAE measure deviation between actual and forecasted values, so the smaller the better the accuracy of forecasts.I find the MAE more practical of course since it isn't squared.

Correct Up Trend (CP)

Correct Down Trend (CD)

CP and CD basically gives the probability of forecast being correct for the next-day USD/JPY direction. So higher is better.

The results compared those of conventional SVR models and SVR network:

USD/JPY	Stat	SVR	SVR N
f(t+1)	sse	47.3925	36.2098
	mae	0.1672	0.1471
	cp	76.8116	77.8468
	cd	79.4118	79.8039
f(t+5)	sse	60.5918	38.4611
	mae	0.1877	0.152
	cp	76.7635	79.668
	cd	78.3465	79.3307
f(t+10)	sse	71.5295	39.7972
	mae	0.2052	0.1544
	cp	77.453	80.167
	cd	78.937	79.9213

So we can see that this is impressive for both SVR and especially the SVR Network. Directionally it's correct close to 80% of the time and an average error of within 0.15 Yen (with respect to daily range), for all three time frame predictions.

A lot of people don't look at this stuff because it takes a lot of thinking, so they convince themselves that it "doesn't work anyway". Well I'm here to tell you that IT DOES.