Business Analytics using R
Very basic time series forecasting technique
- SMA (Simple Moving Average)
- SMA predicts the next value as the average of n past observations
- WMA (Weighted Moving Average)
- WMA assigns linear weights to each past n observations
- EMA (Exponential Moving Average)
- EMA predicts the next observation to be the exponentially weighted mean of past n observations.
Holt Winters forecasting models
This is a time-series forecasting method provided the data is seasonal. Also called as Triple exponential smoothing, Holt winters take into account the level, trend, and seasonality and comprise of three smoothing parameters. Two variations are – Multiplicative and Additive.
AutoRegressive Integrated Moving Average is the generalized version of ARMA models. Its’ equation for a stationary time series is linear in which predictors consist of lags of dependant variable and/or lags of forecast errors.
Generalized AutoRegressive Conditional Heteroskedasticity is used by financial professionals in modeling volatile financial markets. It involves three steps:
- Estimate best fitting autoregressive model
- Compute autocorrelations in the error term
- Test for significance
AutoRegressive Conditional Heteroskedasticity is used for modeling financial time series with time-varying volatility. Most often it’s used where there are short periods of increased variation.
ANNs consists of nodes, which are similar to neurons of the human brain. The neurons are interconnected and they communicate with each other. The nodes have input functions that receive data and operates on the data. The result is passed to other neurons. The output at each node is called its activation or node value
- Feedforward ANN
- Unidirectional flow of information. No feedback loops. Used in pattern recognition
- Feedback ANN
- Feedback loops are present.
Long Short Term Memory networks are RNNs capable of remembering information from long periods of time. It’s’ designed to avoid the long term dependency problem. LSTM is behind many amazing recent achievements of deep learning.
Used for determining the strength of the relationship between a dependant variable and an independent variable. Linear regression uses one independent variable to determine another dependant variable whereas multiple linear regression uses more than one independent variable to explain a dependant variable.