How can reproduce animation for rolling window over time?

Question

I'm experimenting with 1D time-series data and trying to reproduce the following approach via animation over my own data in GoogleColab notebook.

It's about reproducing the animation of STS transformation (implemented by series_to_supervised() function with Lookback steps to past time n_in=9) equal to Backtesting with refit and fixed training size (rolling origin) animation approach introduced skforecast package. It's more about train and test selection over actual time-series data y. Visualize fixed train size and refit and predict next step(s).

ani

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from matplotlib.patches import Rectangle
from matplotlib.animation import FuncAnimation
from IPython.display import HTML
print(pd.__version__)


# Generate univariate (1D) time-series data into pandas DataFrame
import numpy as np
np.random.seed(123) # for reproducibility and get reproducible results

df = pd.DataFrame({
    "TS_24hrs": np.arange(0, 274),
    "count"   : np.abs(np.sin(2 * np.pi * np.arange(0, 274) / 7) + np.random.normal(0, 100.1, size=274)) # generate sesonality
})
#df = pd.read_csv('/content/U2996_24hrs_.csv', header=0, index_col=0).values
print(f"The raw data {df.shape}")
#print(f"The raw data columns {df.columns}")


# visulize data
import matplotlib.pyplot as plt
fig, ax = plt.subplots( figsize=(10,4))

# plot data
df['count'].plot(label=f'data or y', c='red' )
#df['count'].plot(label=f'data',  linestyle='--')
plt.xticks([0, 50, 100, 150, 200, 250, df['TS_24hrs'].iloc[-1]], visible=True, rotation="horizontal")

plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left")
plt.title('Plot of data')
plt.ylabel('count', fontsize=15)
plt.xlabel('Timestamp [24hrs]', fontsize=15)
plt.grid()
plt.show()

# slecet train/test data using series_to_supervised (STS)
from pandas import DataFrame, concat
def series_to_supervised( data, n_in, n_out=1, dropnan=True):
    """
    Frame a time series as a supervised learning dataset.
    Arguments:
        data: Sequence of observations as a list or NumPy array.
        n_in: Number of lag observations as input (X).
        n_out: Number of observations as output (y).
        dropnan: Boolean whether or not to drop rows with NaN values.
    Returns:
        Pandas DataFrame of series framed for supervised learning.
    """
    n_vars = 1 if type(data) is list else data.shape[1]
    df = pd.DataFrame(data)
    cols = list()
    # input sequence (t-n, ... t-1)
    for i in range(n_in, 0, -1):
        cols.append(df.shift(i))
    # forecast sequence (t, t+1, ... t+n)
    for i in range(0, n_out):
        cols.append(df.shift(-i))
    # put it all together
    agg = concat(cols, axis=1)
    # drop rows with NaN values
    if dropnan:
        agg.dropna(inplace=True)
    return agg.values


values=series_to_supervised(df, n_in=9)
data_x,data_y =values[:, :-1], values[:, -1]

print(data_x.shape)
print(data_y.shape)


# define animation function
import matplotlib.pyplot as plt
import pandas as pd
from matplotlib.animation import FuncAnimation

fig, ax = plt.subplots(nrows=1, ncols=1, figsize=(40, 8))
plt.subplots_adjust(bottom=0.25)
plt.xticks(fontsize=12)
ax.set_xticks(range(0, len(data_y), 9))
ax.set_yticks(range(0, 2500, 200))

data_y = pd.Series(data_y)
data_y.plot(color='r', linestyle='-', label="y")
ax.set_title('Time Series')
ax.set_xlabel('Time')
ax.set_ylabel('Value')
ax.legend(loc="upper left")
ax.grid(True, which='both', linestyle='-', linewidth=3)
ax.set_facecolor('gainsboro')

ax.spines['bottom'].set_position('zero')
ax.spines['left'].set_position('zero')
ax.spines['right'].set_color('none')
ax.spines['top'].set_color('none')

nested_list = list(trainX_tss)
lines = [ax.plot([], [], color='g', linestyle='-')[0] for _ in range(len(trainX_tss))]

def init():
    for line in lines:
        line.set_data([], [])
    return lines

def update(frame):
    for i, line in enumerate(lines):
        data = pd.Series(nested_list[i], index=range(frame + i, frame + i + 9))
        line.set_data([], [])
        line.set_data(data.index, data)
    return lines


# define animation setup
anim = FuncAnimation(fig, update,
                    frames=len(nested_list) - 9,
                    init_func=init,
                    interval=500,
                    blit=True,
                    repeat=False)


# Save animation (.gif))
anim.save('BrowniamMotion.gif', writer = "pillow", fps=10 )

# visulize animation in GoogleColab Notebook
# suppress final output
plt.close(0)
HTML(anim.to_html5_video())

STS transformation:

Multi-Step or Sequence Forecasting A different type of forecasting problem is using past observations to forecast a sequence of future observations. This may be called sequence forecasting or multi-step forecasting.

So far, I could reach this output which is so ugly and incorrect.

Is your question just about how to reproduce the _animation_ example at the beginning of your post? I.e. not about the actual training method. — jared, Aug 04 '23 at 14:10
It's still not clear if your goal is just to create the animation or the actual training approach. If it's the training approach, you should focus on that and make the animation afterward. And if it's the animation, why are you sliding the data in your current version rather than a shaded window? — jared, Aug 04 '23 at 21:33
My aim is to create the *animation of the actual training approach* accordingly to be sure that the used training approach is correct. If I cannot reach that, I want to recreate the above animation at the beginning of my post over my generated data. — Mario, Aug 05 '23 at 12:47

score 0 · Accepted Answer · answered Aug 13 '23 at 10:46

In the animation you want, only the green and blue rectangles move and the data is constant. So your update() function should reflect it.

So what you need is to set up initial graph the way you desire, create the parts you want to animate (with init()) and then create frames using update():

import numpy as np
import matplotlib.pyplot as plt
from matplotlib.animation import FuncAnimation
from matplotlib.patches import Rectangle
import pandas as pd
from IPython.display import HTML

# create data
df = pd.DataFrame({
    "TS_24hrs": np.arange(0, 274),
    "count"   : np.abs(np.sin(2 * np.pi * np.arange(0, 274) / 7) + np.random.normal(0, 100.1, size=274)) # generate sesonality
})

# create plot
plt.style.use("ggplot")  # <-- set overall look
fig, ax = plt.subplots( figsize=(10,4))

# plot data
plt.plot(df['TS_24hrs'], df['count'], 'r-', linewidth=0.5,  label='data or y')

# make graph beautiful
plt.plot([], [], 'g-', label="Train", linewidth=8, alpha=0.3) # <-- dummy legend entry
plt.plot([], [], 'b-', label="Test", linewidth=8, alpha=0.3)  # <-- dummy legend entry 
plt.xticks([0, 50, 100, 150, 200, 250, df['TS_24hrs'].iloc[-1]], visible=True, rotation="horizontal")

plt.title('Plot of data')
plt.ylabel('count', fontsize=15)
plt.xlabel('Timestamp [24hrs]', fontsize=15)
plt.grid(True)
plt.legend(loc="upper left")  
fig.tight_layout(pad=1.2)

Y_LIM = 280
TRAIN_WIDTH = 25
TEST_WIDTH = 10

def init():
    rects = [Rectangle((0, 0), TRAIN_WIDTH, Y_LIM, alpha=0.3, facecolor='green'),
             Rectangle((0 + TRAIN_WIDTH, 0), TEST_WIDTH, Y_LIM, alpha=0.3, facecolor='blue')]
    patches = []
    for rect in rects:
            patches.append(ax.add_patch(rect))
    return patches

def update(x_start):
    patches[0].xy = (x_start, 0)
    patches[1].xy = (x_start + TRAIN_WIDTH, 0)    
    return patches


# create "Train" and "Test" areas
patches = init()

ani = FuncAnimation(
    fig,
    update, 
    frames=np.linspace(0, 230, 40),  # all starting points
    interval=500,
    blit=True)


HTML(ani.to_html5_video())

How can reproduce animation for rolling window over time?

1 Answers1