Stationarity and differencing in ML Python - Model Pipeline Trace

Model Pipeline - Stationarity and differencing

This pipeline shows how we prepare time series data by checking if it is stationary and then use differencing to make it stationary. Stationary data means its patterns do not change over time, which helps models learn better.

Data Flow - 4 Stages

1Raw time series data

1000 time points x 1 feature→Collect original time series values→1000 time points x 1 feature

[100, 102, 105, 107, 110, ...]

↓

2Stationarity check

1000 time points x 1 feature→Perform Augmented Dickey-Fuller test to check stationarity→1000 time points x 1 feature + stationarity result

p-value=0.3 (non-stationary)

↓

3Differencing

1000 time points x 1 feature→Subtract previous value from current to remove trend→999 time points x 1 feature

[2, 3, 2, 3, ...] (differences of original series)

↓

4Stationarity re-check

999 time points x 1 feature→Perform Augmented Dickey-Fuller test again→999 time points x 1 feature + stationarity result

p-value=0.01 (stationary)

Training Trace - Epoch by Epoch

Loss
0.5 |****
0.4 |****
0.3 |****
0.2 |****
     1  2  3  4  5 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.45	0.60	Model starts learning with non-stationary data, loss is moderate.
2	0.38	0.68	Loss decreases as model adapts, accuracy improves.
3	0.30	0.75	Better learning after differencing, loss drops further.
4	0.25	0.80	Model converges well on stationary data.
5	0.22	0.83	Training stabilizes with low loss and high accuracy.

Prediction Trace - 4 Layers

Layer 1: Input raw time series value

Layer 2: Differencing operation

Layer 3: Model prediction on differenced data

Layer 4: Inverse differencing

Model Quiz - 3 Questions

Why do we perform differencing on time series data?

A To add noise to the data

B To increase the number of data points

C To make the data stationary by removing trends

D To change the data format to categorical

Key Insight

Making time series data stationary by differencing helps models learn patterns better, leading to lower loss and higher accuracy during training.