Percentile Nearest Rank

[mrjbq7]

Received this issue request:

TA-Lib/ta-lib-python#687

Hello,
I'd like to have similar to TradingView a percentile nearest rank.

I have done it with Numpy and a rolling window until now but its slow compared to anything else in my code and also isn't supported by Numba.

Below is the code so far.

Would be great if this could be added, thanks!

def strided_app(a, L): 
    nrows = a.size-L+1
    n = a.strides[0]
    strided = np.lib.stride_tricks.as_strided(a, shape=(nrows, L), strides=(n, n))
    return strided


# Compute rolling percentiles
perup = np.percentile(strided_app(per_source, window), 60, axis=1, method="nearest")
perdown = np.percentile(strided_app(per_source, window), 45, axis=1, method="nearest")

perup = np.concatenate((np.full(window - 1, np.nan), perup)) 
perdown = np.concatenate((np.full(window - 1, np.nan), perdown))

[sussyboiiii]

I didn't know this existed, thank you for moving my issue over!

[sussyboiiii]

I also have got this faster variant now, which is still rather slow sadly.

@njit(nopython=True, parallel=True, cache=True)
def percentile_nearest(arr, window, percentile):
    n = len(arr)
    result = np.empty(n)
    result[:window-1] = np.nan
    
    for i in prange(window-1, n):
        window_data = arr[i-window+1:i+1]
        sorted_data = np.sort(window_data)
        rank = int(np.floor((percentile/100.0) * (window - 1) + 1)) - 1
        result[i] = sorted_data[rank]
    
    return result