Different Ways to Normalize Time Series

How Scaling Choices Shape Your Models

Time series data has one job: describe how something changes over time. But raw values rarely play nice. Units vary. Magnitudes drift. Outliers throw elbows. Before modeling or comparing signals, you usually need to normalize the data. The trick is choosing the right method.

This guide walks through the most common ways to normalize time series, explains when each one shines, and gives practical Python examples you can drop straight into your workflow.

Note: Unlike previous articles, I’ve chosen to write this one in format that goes straight to the point. I will be showing pros and cons directly with Python code, otherwise, it will take ages if every method needs to have its history explained.

However, I am starting a new series of articles that deal with Quantitative Research where I will present in deeper details the concepts. Stay tuned!

---

Why Normalization Matters

Normalization is more than cosmetics. It affects:

• Model stability

• Training speed

• Similarity comparisons

• Anomaly detection sensitivity

• Feature importance

• Interpretability

For time series, the stakes are even higher because the order of data points matters. A poorly chosen normalization can flatten a trend, exaggerate noise, erase seasonality, or distort what your model believes is important.

Let’s walk through the main options.

Min–Max Scaling - A simple scaling that preserves shape

Min–max scaling transforms your time series so its values sit between 0 and 1 (or any range you choose).

Formula:

x_scaled = (x - min) / (max - min)

Best for

• Models that assume bounded inputs

• Visualization

• Comparing signals with similar ranges

• Neural networks

Weaknesses

• Sensitive to outliers

• Future values outside the observed range break the scale

Python Example

import numpy as np

ts = np.array([12, 15, 14, 20, 18, 25])

min_val = ts.min()
max_val = ts.max()

scaled = (ts - min_val) / (max_val - min_val)
print(scaled)

The Signal Beyond 🚀.

From classic tools that have stood the test of time to fresh innovations like multi-market RSI heatmaps, COT reports, seasonality, and pairs trading recommendation system, the new report is designed to give you a sharper edge in navigating the markets.

Free trial available.

Z-Score Standardization - Centering the data so variation matters more than magnitude

Z-score standardization shifts the series to mean 0 and standard deviation 1.

Formula:

x_scaled = (x - mean) / std

Best for

• Most statistical models

• Anywhere you want variation to matter

• Machine learning pipelines

• Features with different units

Weaknesses

• Outliers distort mean and standard deviation

• If distribution changes over time, scaling must be updated

Python Example

import numpy as np

ts = np.array([12, 15, 14, 20, 18, 25])

mean = ts.mean()
std = ts.std()

z = (ts - mean) / std
print(z)

Robust Scaling - A tougher version of Z-score that ignores outliers

Robust scaling uses the median and interquartile range instead of mean and standard deviation.

Formula:

x_scaled = (x - median) / IQR

Best for

• Data with outliers

• Finance series with spikes

• Sensor readings that sometimes fail

• Real-world messy data

Weaknesses

• If distribution is clean, it can under-represent real variation

Python Example using scikit-learn

import numpy as np
from sklearn.preprocessing import RobustScaler

ts = np.array([12, 15, 14, 20, 18, 25]).reshape(-1, 1)

scaler = RobustScaler()
scaled = scaler.fit_transform(ts)
print(scaled.flatten())

---

Log Scaling - Useful when big values dominate the story

Taking the logarithm compresses large numbers and spreads out small ones.

Formula:

x_scaled = log(x + c)

c is often 1 to avoid log(0).

Best for

• Exponential growth

• Financial time series

• Data with long-tail distributions

• Stabilizing variance

Weaknesses

• Can’t handle negative values without shifts

• Interpretation becomes less intuitive

Python Example

import numpy as np

ts = np.array([5, 8, 12, 200, 350, 800])
log_scaled = np.log(ts + 1)
print(log_scaled)

---

Rolling Normalization - Dynamic scaling that adapts over time

Instead of scaling with global statistics, you normalize within a moving window.

Example:

• Use the last 30 days to compute mean and standard deviation

• Scale today’s value relative to that rolling window

Best for

• Non-stationary time series

• Drift detection

• Online learning

• Signals where “normal” changes over time

Weaknesses

• Computationally heavier

• Results differ based on window size

Python Example

import pandas as pd

ts = pd.Series([12, 14, 16, 30, 28, 26, 40, 42])

rolling_mean = ts.rolling(window=3).mean()
rolling_std = ts.rolling(window=3).std()

z_rolling = (ts - rolling_mean) / rolling_std
print(z_rolling)

Max-Abs Scaling - Scaling for signals that cross zero

Max-Abs scaling keeps the sign and scales values by the largest absolute value.

Formula:

x_scaled = x / max(abs(x))

Best for

• Sparse signals

• Waveforms

• Signed sensor data

• Machine learning models expecting inputs in [-1, 1]

Weaknesses

• Sensitive to a single large value

Python Example

import numpy as np

ts = np.array([-4, -2, 0, 3, 5])
scaled = ts / np.max(np.abs(ts))
print(scaled)

---

Quantile or Rank Transformation - Turning arbitrary shapes into smooth, uniform distributions

This method transforms each value according to its percentile.

Best for

• Strongly skewed data

• When you want identical marginal distributions across features

• Preparing inputs for models sensitive to non-normal shapes

Weaknesses

• Destroys distance relationships

• Flattens patterns if used carelessly

Python Example

import numpy as np
from sklearn.preprocessing import QuantileTransformer

ts = np.array([2, 2, 3, 10, 50, 200]).reshape(-1, 1)

qt = QuantileTransformer(output_distribution=’uniform’)
scaled = qt.fit_transform(ts)
print(scaled.flatten())

Seasonal Normalization - Let each season define its own baseline

If your data has hourly, daily, weekly, or annual patterns, you can normalize separately for each period. Example: normalize all Monday values using Monday statistics.

Best for

• Strong seasonality

• Demand forecasting

• Energy load prediction

• Website traffic patterns

Weaknesses

• Assumes seasonal patterns are stable

• Needs enough data per season

Python Example

import pandas as pd

df = pd.DataFrame({
    “value”: [20, 22, 25, 30, 18, 40, 42],
    “day”: [”Mon”, “Tue”, “Wed”, “Thu”, “Fri”, “Mon”, “Tue”]
})

seasonal = df.groupby(”day”).transform(
    lambda x: (x - x.mean()) / x.std()
)

print(seasonal)

Normalization is a small step that changes everything downstream. The wrong method hides patterns. The right one reveals them. The best normalization is the one that respects the story your data is trying to tell.

Do you want to master Deep Learning techniques tailored for time series, trading, and market analysis🔥? My book breaks it all down from basic machine learning to complex multi-period LSTM forecasting while going through concepts such as fractional differentiation and forecasting thresholds. Get your copy here 📖!