Time Series: Dependence, Trend, Seasonality & Forecasting

Statistics
Time Series
Forecasting
Data Science
Learn why time breaks ordinary i.i.d. assumptions and connect decomposition, autocorrelation, backtesting and forecasting to practical ML.
Author

MLStation

Published

October 1, 2026

Most introductory statistics quietly assumes observations can be treated as independent.

Time series often refuses to cooperate.

Yesterday can predict today. December can resemble last December. A trend can make the future unlike the past average.

TipWhat you will learn
  • Trend, seasonality, noise and autocorrelation
  • Lagged relationships
  • Why random train/test splitting can leak the future
  • Rolling-origin backtesting
  • How statistical forecasting connects to ML

A time-series decomposition

A useful conceptual model is

Y_t=T_t+S_t+R_t,

where:

  • T_t = trend,
  • S_t = seasonal component,
  • R_t = remainder/noise.

Interactive: build a signal

Change one component at a time. The observed line is the mixture.

Autocorrelation

Autocorrelation at lag k compares Y_t with Y_{t-k}.

A strong lag-7 relationship in daily data may indicate weekly seasonality.

Show the code 
import numpy as np, pandas as pd
import plotly.graph_objects as go

rng=np.random.default_rng(18)
t=np.arange(365)
y=20+.02*t+3*np.sin(2*np.pi*t/7)+rng.normal(0,1.2,len(t))

lags=np.arange(1,31)
acf=[np.corrcoef(y[k:],y[:-k])[0,1] for k in lags]

fig=go.Figure(go.Bar(x=lags,y=acf))
fig.update_layout(height=340,xaxis_title="Lag",yaxis_title="Autocorrelation")
fig.show()

The cardinal rule: do not train on the future

Random splitting can create leakage:

BAD FOR MANY TIME-SERIES TASKS
past + future mixed → random train/test

Prefer chronological evaluation:

TRAIN                 TEST
───────────────|────────────
past                     future

Animation: rolling-origin backtesting

Show the code 
import numpy as np, pandas as pd
import plotly.express as px

rng=np.random.default_rng(19)
t=np.arange(160)
y=10+.04*t+2*np.sin(2*np.pi*t/20)+rng.normal(0,.7,len(t))
rows=[]
for cutoff in [60,80,100,120,140]:
    for i,(tt,yy) in enumerate(zip(t,y)):
        part="train" if tt<=cutoff else "future"
        rows.append({"cutoff":cutoff,"time":tt,"y":yy,"part":part})
df=pd.DataFrame(rows)
fig=px.scatter(df,x="time",y="y",color="part",animation_frame="cutoff",
               range_x=[0,160],range_y=[5,20])
fig.update_layout(height=380)
fig.show()

Each frame moves the training frontier forward. This is much closer to how forecasting models are used in reality.

DS/ML Bridge

Statistical forecasting and ML share the same core evaluation discipline:

features from past only
      ↓
fit model
      ↓
predict unseen future
      ↓
advance time
      ↓
repeat

ML models may use lag features, rolling statistics, calendar features and exogenous variables. Statistical models may encode trend, autoregression and seasonality directly.

The algorithm matters—but temporal validation design matters first.

WarningLeakage can make a bad forecast look excellent

Any feature computed using information unavailable at prediction time can invalidate evaluation, even if the code runs perfectly.

What should you save?

forecast_run/
├── data_cutoff.txt
├── feature_config.json
├── split_definition.json
├── model_config.json
├── forecasts.csv
└── metrics.json

A forecast is inseparable from the time at which it was made.

Summary

  • Time series contain ordered dependence.
  • Trend, seasonality and autocorrelation create structure across time.
  • Random splitting can leak future information.
  • Rolling-origin evaluation mimics real forecasting.
  • Statistical and ML forecasting differ in machinery but share the need for leakage-safe temporal validation.

Next: Revisit the series as a connected toolkit: describe → model uncertainty → infer → experiment → predict → validate.

Back to top