使用日期偏移不正確的scikit-learn線性模型預測

Question

我試圖預測時間序列數據，但是通過在訓練和預測之前抵消了date_offset時間點的結果。這樣做的原因是嘗試使用當前數據預測未來的date_offset時間點。一個例子見http://glowingpython.blogspot.co.za/2015/01/forecasting-beer-consumption-with.html。

因此，在總結： data = [1,2,3,4,5]應該預測result = [2,3,4,5,6]如果date_offset = 1

下面對劇情結果表明紅線由date_offset被轉移，而不是預測date_offset到未來。無論我有多大date_offset，它都在不斷變化，並且不會預測我的最後結果，即result = 5（已知）。實際上，紅線根本不應該移位，只是寬鬆的準確度更大的date_offset成爲。我究竟做錯了什麼？

參見下面的例子代碼和生成的圖像：

from sklearn import linear_model 
import matplotlib.pyplot as plt 
import numpy as np 

date_offset = 1 

data = np.array([9330.0, 9470.0, 9550.0, 9620.0, 9600.0, 9585.0, 9600.0, 9600.0, 9430.0, 9460.0, 9450.0, 9650.0, 9620.0, 9650.0, 9500.0, 9400.0, 9165.0, 9100.0, 8755.0, 8850.0, 8990.0, 9150.0, 9195.0, 9175.0, 9250.0, 9200.0, 9350.0, 9280.0, 9370.0, 9470.0, 9445.0, 9440.0, 9280.0, 9325.0, 9170.0, 9270.0, 9200.0, 9450.0, 9510.0, 9371.0, 9499.0, 9499.0, 9400.0, 9500.0, 9550.0, 9670.0, 9700.0, 9760.0, 9767.4599999999991, 9652.0, 9520.0, 9600.0, 9610.0, 9700.0, 9825.0, 9900.0, 9950.0, 9801.0, 9770.0, 9545.0, 9630.0, 9710.0, 9700.0, 9700.0, 9600.0, 9615.0, 9575.0, 9500.0, 9600.0, 9480.0, 9565.0, 9510.0, 9475.0, 9600.0, 9400.0, 9400.0, 9400.0, 9300.0, 9430.0, 9410.0, 9380.0, 9320.0, 9000.0, 9100.0, 9000.0, 9200.0, 9210.0, 9251.0, 9460.0, 9400.0, 9600.0, 9621.0, 9440.0, 9490.0, 9675.0, 9850.0, 9680.0, 10100.0, 9900.0, 10100.0, 9949.0, 10040.0, 10050.0, 10200.0, 10400.0, 10350.0, 10200.0, 10175.0, 10001.0, 10110.0, 10400.0, 10401.0, 10300.0, 10548.0, 10515.0, 10475.0, 10200.0, 10481.0, 10500.0, 10540.0, 10559.0, 10300.0, 10400.0, 10202.0, 10330.0, 10450.0, 10540.0, 10540.0, 10650.0, 10450.0, 10550.0, 10501.0, 10206.0, 10250.0, 10345.0, 10225.0, 10330.0, 10506.0, 11401.0, 11245.0, 11360.0, 11549.0, 11415.0, 11450.0, 11460.0, 11600.0, 11530.0, 11450.0, 11402.0, 11299.0]) 
data = data[np.newaxis].T 

results = np.array([9470.0, 9545.0, 9635.0, 9640.0, 9600.0, 9622.0, 9555.0, 9429.0, 9495.0, 9489.0, 9630.0, 9612.0, 9630.0, 9501.0, 9372.0, 9165.0, 9024.0, 8780.0, 8800.0, 8937.0, 9051.0, 9100.0, 9166.0, 9220.0, 9214.0, 9240.0, 9254.0, 9400.0, 9450.0, 9470.0, 9445.0, 9301.0, 9316.0, 9170.0, 9270.0, 9251.0, 9422.0, 9466.0, 9373.0, 9440.0, 9415.0, 9410.0, 9500.0, 9520.0, 9620.0, 9705.0, 9760.0, 9765.0, 9651.0, 9520.0, 9600.0, 9610.0, 9700.0, 9805.0, 9900.0, 9950.0, 9800.0, 9765.0, 9602.0, 9630.0, 9790.0, 9710.0, 9800.0, 9649.0, 9580.0, 9780.0, 9560.0, 9501.0, 9511.0, 9530.0, 9498.0, 9475.0, 9595.0, 9500.0, 9460.0, 9400.0, 9310.0, 9382.0, 9375.0, 9385.0, 9320.0, 9100.0, 8990.0, 9045.0, 9129.0, 9201.0, 9251.0, 9424.0, 9440.0, 9500.0, 9621.0, 9490.0, 9512.0, 9599.0, 9819.0, 9684.0, 10025.0, 9984.0, 10110.0, 9950.0, 10048.0, 10095.0, 10200.0, 10338.0, 10315.0, 10200.0, 10166.0, 10095.0, 10110.0, 10400.0, 10445.0, 10360.0, 10548.0, 10510.0, 10480.0, 10180.0, 10488.0, 10520.0, 10510.0, 10565.0, 10450.0, 10400.0, 10240.0, 10338.0, 10410.0, 10540.0, 10481.0, 10521.0, 10530.0, 10325.0, 10510.0, 10446.0, 10249.0, 10236.0, 10211.0, 10340.0, 10394.0, 11370.0, 11250.0, 11306.0, 11368.0, 11415.0, 11400.0, 11452.0, 11509.0, 11500.0, 11455.0, 11400.0, 11300.0, 11369.0]) 

# Date offset to predict next i-days results 
data = data[:-date_offset] 
results = results[date_offset:] 

train_data = data[:-50] 
train_results = results[:-50] 

test_data = data[-50:] 
test_results = results[-50:] 

regressor = linear_model.BayesianRidge(normalize=True) 
regressor.fit(train_data, train_results) 

plt.figure(figsize=(8,6)) 
plt.plot(regressor.predict(test_data), '--', color='#EB3737', linewidth=2, label='Prediction') 
plt.plot(test_results, label='True', color='green', linewidth=2) 
plt.legend(loc='best') 
plt.show() 

所有的

Answer 1

首先，該模型是不是真的壞。例如，當真實值爲10450時，它預測10350，這真的很接近。而且，顯然，預測點越晚，預測越不準確，因爲方差在增加，有時甚至偏差也在增長。你不能期待相反的情況。其次，它是一個線性模型，所以當預測變量本質上不是線性時，它不可能是絕對精確的。

第三，必須小心選擇一個預測變量。例如，在這種情況下，您可能會嘗試不預測時間T的值，而是預測時間T的值的變化（即C [T] = V [T] -V [T-1]）或移動平均值最後的K值。在這裏，你可能（或者相反，可能不會）發現你正試圖對所謂的「隨機遊走」進行建模，這種隨機遊走很難通過其隨機性來準確預測。最後，您可能會考慮其他模型，如ARIMA，它們更適合預測時間序列。

Answer 2

添加回organize_data步：

import matplotlib.pyplot as plt 
import numpy as np 
import pandas as pd 
from sklearn import linear_model 
def organize_data(to_forecast, window, horizon): 
    """ 
    Input: 
     to_forecast, univariate time series organized as numpy array 
     window, number of items to use in the forecast window 
     horizon, horizon of the forecast 
    Output: 
     X, a matrix where each row contains a forecast window 
     y, the target values for each row of X 
    """ 
    shape = to_forecast.shape[:-1] + \ 
      (to_forecast.shape[-1] - window + 1, window) 
    strides = to_forecast.strides + (to_forecast.strides[-1],) 
    X = np.lib.stride_tricks.as_strided(to_forecast, 
             shape=shape, 
             strides=strides) 
    y = np.array([X[i+horizon][-1] for i in range(len(X)-horizon)]) 
    return X[:-horizon], y 

data = np.array([9330.0, 9470.0, 9550.0, 9620.0, 9600.0, 9585.0, 9600.0, 9600.0, 9430.0, 9460.0, 9450.0, 9650.0, 9620.0, 9650.0, 9500.0, 9400.0, 9165.0, 9100.0, 8755.0, 8850.0, 8990.0, 9150.0, 9195.0, 9175.0, 9250.0, 9200.0, 9350.0, 9280.0, 9370.0, 9470.0, 9445.0, 9440.0, 9280.0, 9325.0, 9170.0, 9270.0, 9200.0, 9450.0, 9510.0, 9371.0, 9499.0, 9499.0, 9400.0, 9500.0, 9550.0, 9670.0, 9700.0, 9760.0, 9767.4599999999991, 9652.0, 9520.0, 9600.0, 9610.0, 9700.0, 9825.0, 9900.0, 9950.0, 9801.0, 9770.0, 9545.0, 9630.0, 9710.0, 9700.0, 9700.0, 9600.0, 9615.0, 9575.0, 9500.0, 9600.0, 9480.0, 9565.0, 9510.0, 9475.0, 9600.0, 9400.0, 9400.0, 9400.0, 9300.0, 9430.0, 9410.0, 9380.0, 9320.0, 9000.0, 9100.0, 9000.0, 9200.0, 9210.0, 9251.0, 9460.0, 9400.0, 9600.0, 9621.0, 9440.0, 9490.0, 9675.0, 9850.0, 9680.0, 10100.0, 9900.0, 10100.0, 9949.0, 10040.0, 10050.0, 10200.0, 10400.0, 10350.0, 10200.0, 10175.0, 10001.0, 10110.0, 10400.0, 10401.0, 10300.0, 10548.0, 10515.0, 10475.0, 10200.0, 10481.0, 10500.0, 10540.0, 10559.0, 10300.0, 10400.0, 10202.0, 10330.0, 10450.0, 10540.0, 10540.0, 10650.0, 10450.0, 10550.0, 10501.0, 10206.0, 10250.0, 10345.0, 10225.0, 10330.0, 10506.0, 11401.0, 11245.0, 11360.0, 11549.0, 11415.0, 11450.0, 11460.0, 11600.0, 11530.0, 11450.0, 11402.0, 11299.0]) 

train_window = 50 
k = 5 # number of previous observations to use 
h = 2 # forecast horizon 
X,y = organize_data(data, k, h) 

train_data = X[:train_window] 
train_results = y[:train_window] 

test_data = X[train_window:] 
test_results = y[train_window:] 

regressor = linear_model.BayesianRidge(normalize=True) 
regressor.fit(train_data, train_results) 

plt.figure(figsize=(8,6)) 
plt.plot(regressor.predict(X), '--', color='#EB3737', linewidth=2, label='Prediction') 
plt.plot(y, label='True', color='green', linewidth=2) 
plt.legend(loc='best') 
plt.show()