Time Series — USTHB and EViews-style methodology

This page is the course-oriented reference for the applied time-series layer. It describes the methodology and reasoning behind the Python workflow; the Time Series User Guide provides the practical module-by-module instructions.

1. From raw data to a model

The workflow is deliberately sequential:

raw series
→ workfile / sample
→ descriptive analysis
→ transformations
→ stationarity
→ ACF/PACF
→ model specification
→ estimation
→ residual diagnostics
→ selection
→ forecast

This is close to the workflow EViews users already know, but the full analysis is now a Python script that can be versioned and rerun.

2. Workfile thinking

from stochx.timeseries import Workfile

wf = Workfile.from_csv(
    "macro.csv",
    date_column="DATE",
    frequency="M",
)
wf.set_sample("2010-01-01 2024-12-01")

A workfile contains named series and a current estimation sample. This is the central object for multi-series empirical analysis.

3. Expressions and transformations

The expression layer keeps common EViews conventions familiar:

wf.eval("GDP(-1)")
wf.eval("D(GDP)")
wf.eval("DLOG(CPI)")
wf.eval("LOG(IP)")
wf.generate("GDP_GROWTH", "DLOG(GDP)")

Use transformations to express a modeling assumption explicitly: remove trend/level behavior, stabilize variance, construct growth rates or create lagged regressors.

4. Stationarity

A unit-root decision must identify the deterministic specification first. StochX exposes DF/ADF, KPSS and Phillips–Perron procedures.

from stochx.timeseries import adf

result = adf(wf["GDP"], regression="c", lags=1, autolag=None)
print(result.summary())
print(result.interpret())

The test output should be read together with the series plot and economic context.

5. Identification with ACF/PACF

Once a working stationary series has been selected, inspect the autocorrelation and partial autocorrelation functions.

from stochx.timeseries import correlogram

corr = correlogram(wf["GDP_GROWTH"].values, nlags=24)
print(corr.table())

Use ACF/PACF as model-identification evidence, not as a mechanical order generator.

6. Model families

StochX covers:

• AR(p);
• MA(q);
• ARMA(p,q);
• ARIMA(p,d,q);
• SARIMA(p,d,q) × (P,D,Q,s);
• regression equations with ARMA error terms.

For example:

from stochx.timeseries import fit_arima

result = fit_arima(wf["GDP"], order=(1, 1, 1))
print(result.summary())

7. Box–Jenkins

For systematic candidate comparison, use the explicit workflow:

from stochx.timeseries import (
    identify_box_jenkins,
    estimate_box_jenkins_candidates,
    validate_box_jenkins_candidates,
    select_box_jenkins_model,
    forecast_box_jenkins,
)

ident = identify_box_jenkins(wf["GDP"], d=1, nlags=24, max_p=3, max_q=3)
estimation = estimate_box_jenkins_candidates(wf["GDP"], ident.candidate_orders)
validation = validate_box_jenkins_candidates(estimation, lags=12, alpha=0.05)
selection = select_box_jenkins_model(validation, criterion="aic")
forecast = forecast_box_jenkins(selection, steps=12, alpha=0.05)

The important distinction is that estimation success is not model adequacy. Residual validation is a separate decision.

8. Regression equations and ARMA errors

EViews-style equation specifications remain textual:

eq = wf.ls(
    "TBILL C LOG(M1) LOG(CPI) LOG(IP) @TREND AR(1) MA(1)",
    name="EQ20",
)

C is the intercept, @TREND is the deterministic trend term where supported, and AR(n) / MA(n) refer to the disturbance process.

9. Diagnostics

Residual validation includes serial-correlation, normality and variance diagnostics where the selected model supports them. Typical tools include Ljung–Box, Breusch–Godfrey, Jarque–Bera, Kolmogorov–Smirnov, Breusch–Pagan and ARCH tests.

A report should state the null hypothesis, statistic, p-value and modeling consequence.

10. Forecasting

Forecast reports should include point forecasts and uncertainty. StochX keeps standard errors, lower/upper bounds, horizon and future index together when supported.

11. State-space extension

The linear-Gaussian state-space layer provides an alternative representation when latent states, missing observations or recursive dynamics matter.

The workflow is:

model → Kalman filter → RTS smoother → innovation diagnostics → adequacy → forecast

See the state-space guide.

12. EViews migration principle

The migration target is not a GUI imitation. The target is methodological familiarity:

• the same sequence of statistical decisions;
• recognizable names and expressions;
• table-oriented results;
• explicit estimation samples;
• reproducible scripts;
• documented numerical conventions.

Where exact EViews benchmark outputs are available in the repository, they are treated as frozen numerical fixtures rather than informal claims of universal parity.