Penalized Least Squares for Smoothing Financial Time Series

Modeling of financial time series data by methods of artificial intelligence is difficult because of the extremely noisy nature of the data. A common and simple form of filter to reduce the noise originated in signal processing, the finite impulse response (FIR) filter. There are several of these noise reduction methods used throughout the financial instrument trading community. The major issue with these filters is the delay between the filtered data and the noisy data. This delay only increases as more noise reduction is desired. In the present marketplace, where investors are competing for quality and timely information, this delay can be a hindrance. This paper proposes a new FIR filter derived with the aim of maximizing the level of noise reduction and minimizing the delay. The model is modified from the old problem of time series graduation by penalized least squares. Comparison between five different methods has been done and experiment results have shown that our method is significantly superior to the alternatives in both delay and smoothness over short and middle range delay periods.