L28 Hurst Exponent Estimation under ARFIMA
- Fractionally Integrated \(\implies\) \(d\) can take fractional values
Hurst Exponent - An Index of Long-Range Dependence¶
- \(H\) can be used as a measure of long-term memory of TS.
- Developed in hydrology → optimum dam size (river Nile)
- Used in finance, hydrology and physics (behavior of physics over time)
Construction¶
We need some ingredients…
\(y_{t}\) for \(t=1,2,\dots,n\) be a realization with mean \(\bar{y}_{n}\) and variance \(S_{n}^{2}\)
- Mean adjusted partial sums defined as,
Example, \(Z_{1} = y_{1} - \bar{y}_{1}\), \(Z_{2} = y_{1} + y_{2} - \bar{y}_{t} - 2\bar{y}_{t}\) etc.
- Adjusted range is defined as
- Rescaled adjusted range = \(\dfrac{R_{n}}{S_{n}}\)
- \(E(\dfrac{R_{n}}{S_{n}}) \propto Cn^H\) as \(n\to \infty\)
- If the realization is i.i.d., \(H = 0.5\)
For large \(n\),
$$
\log\left(\dfrac{R_{n}}{S_{n}}\right) = \alpha + H \times \log(n)
$$
- Can be viewed as \(\log\left(\dfrac{R_{n}}{S_{n}}\right)\) regressed on \(\log(n)\)
- This is how we can estimate, \(\alpha\) and \(H\)
Properties of \(H\)¶
- \(H \in (0.5, 1):\) Long memory structure
- \(H \in [1,\infty):\) infinite variance, non-stationary
- \(H \in (0,0.5):\) Anti-persistence structure exists
- \(H = 0.5:\) White noise structure exists (i.i.d.)
Application of \(H\) exponent¶
- Calculate Hurst exponent, to gain insights into the nature of price movements.
- Stock price shows a persistent trend, trending market or momentum
- Investors can use this persistence ka information to apply a trend-following strategies \(\implies\) price movements will persist
- Stock price follows random walk \(\implies\) no observable trend. Equally likely to increase or decrease
- Stock price shows mean-reverting behavior.
- After an increase, more likely to decrease and vice versa. (Stock overbought and oversold)