Stock Index Predictions with Machine Learning

The NIKKEI225 empirical tests from "Forecasting stock market direction with support vector machine" is pretty cool. The document also contains a very basic, easy to follow explanation of SVM theory.

They tested a 1-week directional forecast accuracy against some existing quantitative means:
RW (Random Walk)
LDA (Linear Discriminant Analysis)
QDA (Quadratic Discriminant Analysis)
EBNN (Elman Backpropogation Neural Network)

NIKKEI 225 Forecasting model:

The tested data ranges from Jan. 1, 1990 to Dec. 31, 2002, a total of 676 weeks tested.

Input variables:
t= week of directional forecast, so
(t-1)= 1 week prior to t

1) S&P(t-1): Value of US S&P500 at (t-1)
2) jpy(t-1): Value of JPY/USD at (t-1),
3) N(t-1): Value of NIKKEI225 at (t-1)

Output:
NIKKEI225 1-week Direction Forecast:
F(t)= [S&P(t-1), jpy(t-1)]

Accuracy results:

RW- 50% (we all know the market isn't random by now right?)
LDA- 55% (the market does not move linearly)
QDA- 69% (predictions become dramatically better once we assume nonlinearity)
EBNN- 69% (Neural Networks indeed offer an edge apparently)
SVM- 73%

According to Wang, Nakamori, and Huang, SVM has a critical advantage aginst other methods, Structural Risk Minization. SVMs overcome problems from "overfitting" since most other machine learning methods only seek to minimize Empirical Risk.

Pretty cool huh!