构建列间含指定间隔的日度数据滚动窗口矩阵技术问询

Got it, let's break down how to solve this problem clearly. First, let's align on the core requirements with your example to make sure we're on the same page:

Problem Recap #

You have a daily time series vector (21 years of data) and want to build a matrix where:

• Each column represents a 365-day rolling window
• The next window starts 30 days later than the previous one (so windows overlap by 365-30=335 days)
• The matrix is structured such that each row corresponds to the same position in every window (e.g., row 1 = first day of each window, row 2 = second day, etc.)

Your small example confirms this: with vec = [1,2,...,17], window_size=5, and a 3-step start interval, we get the matrix you provided where each column is a shifted window, and rows align the same position across windows.

Solution 1: Python (NumPy/Pandas) #

NumPy is perfect for efficient matrix operations here. Here's a reusable function, plus a test with your example:

Code Implementation #

import numpy as np
import pandas as pd

def build_rolling_matrix(vec, window_size, step_size):
    # Calculate all valid starting indices (no out-of-bounds windows)
    start_indices = np.arange(0, len(vec) - window_size + 1, step_size)
    # Initialize matrix: rows = window size, columns = number of valid windows
    rolling_matrix = np.zeros((window_size, len(start_indices)))
    # Fill each column with the corresponding window data
    for col_idx, start in enumerate(start_indices):
        rolling_matrix[:, col_idx] = vec[start:start+window_size]
    return rolling_matrix

# Test with your example
test_vec = np.array([1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17])
window_size = 5
step_size = 3  # This is your "30-day interval" for daily data

result = build_rolling_matrix(test_vec, window_size, step_size)
print(result)

Output #

[[ 1.  4.  7. 10. 13.]
 [ 2.  5.  8. 11. 14.]
 [ 3.  6.  9. 12. 15.]
 [ 4.  7. 10. 13. 16.]
 [ 5.  8. 11. 14. 17.]]

For Your Daily Data Use Case #

Just plug in your parameters:

• window_size=365
• step_size=30
• If your data has date labels (e.g., a Pandas Series), you can verify window dates like this:

# Example: Generate simulated daily data with dates
date_range = pd.date_range(start="2000-01-01", end="2018-01-01", freq="D")
daily_data = np.random.randn(len(date_range))  # Replace with your actual data

# Build the matrix
final_matrix = build_rolling_matrix(daily_data, 365, 30)

# Check window date ranges
start_dates = date_range[::30][:final_matrix.shape[1]]
end_dates = start_dates + pd.Timedelta(days=364)  # 365-day window = start to start+364

for idx, (start, end) in enumerate(zip(start_dates, end_dates)):
    print(f"Window {idx+1}: {start.date()} to {end.date()}")

Solution 2: R #

If you prefer R, here's an equivalent implementation using base R functions:

Code Implementation #

build_rolling_matrix <- function(vec, window_size, step_size) {
  # Calculate valid starting indices
  start_indices <- seq(from = 1, to = length(vec) - window_size + 1, by = step_size)
  # Extract each window and transpose to match your desired structure
  rolling_matrix <- t(sapply(start_indices, function(start) {
    vec[start:(start + window_size - 1)]
  }))
  return(rolling_matrix)
}

# Test with your example
test_vec <- c(1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17)
window_size <- 5
step_size <- 3

result <- build_rolling_matrix(test_vec, window_size, step_size)
print(result)

Output #

[,1] [,2] [,3] [,4] [,5]
[1,]    1    4    7   10   13
[2,]    2    5    8   11   14
[3,]    3    6    9   12   15
[4,]    4    7   10   13   16
[5,]    5    8   11   14   17

Key Note on n_interval #

You mentioned n_interval is the difference between the first point of the next window and the last point of the previous window. For our code, this value is equal to step_size - window_size:

• In your example: 3 - 5 = -1 (matches the next window's first point (4) minus previous window's last point (5) = -1)
• For your daily data: 30 - 365 = -335 (which makes sense—each new window starts 30 days after the prior start, so it overlaps 335 days with the previous window)

内容的提问来源于stack exchange，提问作者mk_sch