Does ARIMA need stationary data?

Yes. The 'I' in ARIMA stands for integrated. After d-th differences the series should be approximately stationary in mean and variance. The ADF test, KPSS test, and a flat ACF tail are common checks.

What is AIC and BIC, and which is better?

AIC (Akaike Information Criterion) and BIC (Bayesian Information Criterion) are model-fit penalties — lower is better. BIC penalises complexity more strongly than AIC, so for small samples or parsimony BIC is preferred; for forecasting AIC often performs well.

What is the Ljung-Box test in ARIMA?

The Ljung-Box test checks whether residuals are white noise. A non-significant p-value (p > 0.05) indicates no remaining autocorrelation, meaning the ARIMA model has captured the time-dependent structure adequately.

How do I report ARIMA results in a research paper?

Report the model order ARIMA(p,d,q), parameter estimates with standard errors, log-likelihood, AIC and BIC, residual diagnostics (Ljung-Box Q, p-value), and out-of-sample forecast accuracy (RMSE, MAPE) where applicable.

ARIMA Model Calculator – Free Online Time Series Forecasting Tool

📥 1. Enter Your Time Series

Time series values (comma-separated):

Format: 52, 48, 55, 61, 47, ... — values separated by commas (newlines, tabs and semicolons also accepted).

Use sample dataset:

Upload CSV or Excel file:

Supports .csv .txt .xlsx .xls. Click columns to load each as a separate cluster (group).

Add multiple time-series clusters/groups. Each cluster gets fit independently. The group name is editable.

⚙️ Model settings

AR order (p)

Difference (d)

MA order (q)

Forecast horizon (h)

Significance α

Auto order (AIC)

📊 2. Model Results

Results will appear here after you click Run ARIMA Analysis.

🧠 3. Interpretation Results & How to Write Your Results

📖 Detailed Interpretation Results

▶ Run the analysis above to populate the interpretation with your computed values.

✍️ How to Write Your Results in Research (5 Examples)

▶ Run the analysis above to auto-fill all five reporting examples with your numbers.

🏁 4. Conclusion

▶ Run the analysis to generate the full study conclusion.

🧮 5. Technical Notes & Formulas

A. Formulas

ARIMA(p, d, q) model. Let y_t be the original series and let ∇^dy_t be the d-th difference. Then:

∇^d y_t = c + φ₁·∇^d y_{t-1} + … + φ_p·∇^d y_{t-p} + θ₁·ε_{t-1} + … + θ_q·ε_{t-q} + ε_t

y_t — observed value at time t
∇^dy_t — d-th difference, e.g. ∇y_t = y_t − y_t−1
c — drift / intercept term
φ_i — autoregressive (AR) coefficient at lag i
θ_j — moving-average (MA) coefficient at lag j
ε_t ~ WN(0, σ²) — white-noise error

ACF at lag k:

ρ_k = Σ_{t=k+1..n} (y_t − ȳ)(y_{t-k} − ȳ) ───────────────────────────────────── Σ_{t=1..n} (y_t − ȳ)²

PACF at lag k is the last coefficient φ_kk from the Yule–Walker / Durbin–Levinson recursion fitting an AR(k) model to the series.

AIC and BIC for the fitted model:

AIC = 2k − 2·ln(L) BIC = k·ln(n) − 2·ln(L) where k = p + q + (1 if intercept else 0) + 1 (for σ²)

Ljung–Box Q statistic on residuals (lag h):

Q = n(n+2) · Σ_{k=1..h} ρ̂_k² / (n − k) Under H₀ (white noise): Q ~ χ²(h − p − q)

Forecast variance. For an ARIMA(p, d, q) forecast at horizon h, the variance grows with h via the MA(∞) representation: σ²_h = σ² · Σ_j=0..h−1 ψ_j². The 95% prediction interval is ŷ_n+h ± 1.96·√σ²_h.

B. Notes

ADF test should reject at α = 0.05 after differencing (i.e. the d-th differenced series should be stationary).
Residuals should be white noise: a non-significant Ljung–Box result is desired.
Compare AIC and BIC across competing (p, d, q) candidates — lower is better.
This calculator uses an OLS-style approximation to the Box–Jenkins MLE for the AR and MA coefficients, suitable for educational and exploratory use.

📌 6. When to Use the ARIMA Model

ARIMA is appropriate when your data are an evenly spaced time series, show autocorrelation, and you want short- to medium-horizon forecasts that account for past values and past forecast errors.

Decision checklist

✅ Univariate time series with regular intervals
✅ At least ~50 observations
✅ Series can be made stationary by differencing (d = 0, 1, or 2)
✅ ACF / PACF show interpretable autocorrelation structure
❌ Strong seasonal pattern → use SARIMA(p,d,q)(P,D,Q)_s instead
❌ Multiple correlated series → consider VAR
❌ Non-linear or volatility-driven dynamics → consider ARCH / GARCH or state-space models

Real-world examples

Economics — forecasting quarterly GDP growth or monthly inflation rates.
Retail / Sales — projecting next quarter's revenue from monthly sales history.
Energy — predicting daily electricity demand or wind-power output.
Public health / Ecology — modelling weekly disease incidence or wildlife counts.

Sample-size guidance

Minimum ~50 observations for a stable ARIMA(p,d,q) fit.
Forecast horizon should not exceed ~25% of the historical length.
Use cross-validation or hold-out evaluation (RMSE, MAPE) for honest performance estimates.

📘 7. How to Use This Tool — Step by Step

Enter your data — paste comma-separated values (default), upload a CSV/Excel, or add multiple clusters in the Multi-series tab.
Choose a sample dataset — five built-in datasets cover trend, seasonality, finance, traffic, and economics.
Set the orders (p, d, q) — start with d from the differencing required for stationarity; pick p from the PACF cut-off, q from the ACF cut-off.
Or use Auto order — the tool searches up to (3,2,3) and picks the model with the lowest AIC.
Click Run ARIMA Analysis — model fits in the browser, no server, no upload of your data.
Read the model summary cards — model order, log-likelihood, AIC, BIC, residual SD, Ljung-Box p-value.
Inspect the four plots — series + forecast, ACF, PACF, residuals.
Check residuals — Ljung-Box p > 0.05 indicates the model has captured the autocorrelation structure.
Read the interpretation block — auto-filled paragraphs translate the numbers into plain English.
Export — Download Doc (text) or Download PDF (print-ready report).

❓ 8. Frequently Asked Questions

Q1. What is the ARIMA model and when should I use it?

ARIMA — AutoRegressive Integrated Moving Average — is a classical time-series model that combines past values (AR), differencing for stationarity (I), and past forecast errors (MA). Use it when your data are evenly spaced, show autocorrelation, and you want short- to medium-term forecasts with reasonable interpretability.

Q2. What do p, d, and q mean in ARIMA(p,d,q)?

p is the AR order — how many past values feed into today's value. d is the number of times you differenced the series to make it stationary. q is the MA order — how many past forecast errors enter the model. A series like ARIMA(1,1,1) uses one AR term, one differencing, and one MA term.

Q3. How do I choose the right (p, d, q)?

Choose d so the d-th differenced series looks stationary (flat mean, constant variance, ADF rejects the unit root). Then use the ACF and PACF: a sharp PACF cut-off at lag p suggests AR(p); a sharp ACF cut-off at lag q suggests MA(q). Compare candidates by AIC/BIC and verify residuals are white noise (Ljung–Box).

Q4. What does the p-value mean for the Ljung–Box test?

The Ljung–Box test asks: are the residuals white noise? A high p-value (typically p > 0.05) means we cannot reject "no autocorrelation in residuals" — the model has adequately captured the temporal structure. A small p-value warns that something is still left in the residuals.

Q5. Does my time series need to be stationary?

The differenced series — that is, ∇^dy_t — must be approximately stationary in mean and variance. Trend or non-constant variance violates ARMA's assumptions, which is why the "I" (integration / differencing) step exists. Use ADF or KPSS to check, and apply log/Box–Cox transforms when variance grows with the level.

Q6. What is the difference between ARIMA and SARIMA?

ARIMA models non-seasonal patterns. SARIMA adds seasonal terms (P, D, Q) at period s — for example monthly data with yearly seasonality is often modelled as SARIMA(1,1,1)(1,1,1)₁₂. If your ACF shows large spikes at lag 12 or 24, you probably need SARIMA, not plain ARIMA.

Q7. What is AIC vs BIC, and which should I trust?

Both penalise model complexity, but BIC penalises more strongly. For pure forecasting accuracy, AIC often performs slightly better; for parsimony and small samples, BIC is preferred. In practice, report both — and confirm with residual diagnostics and out-of-sample MAPE/RMSE.

Q8. How many data points do I need for ARIMA?

At least 50 observations for a stable fit, more for higher-order models. For seasonal data (SARIMA), aim for 4–5 full seasonal cycles. Forecast horizon should be much smaller than the historical length — forecasting 24 months from 30 months of history is unwise.

Q9. How do I report ARIMA results in a paper?

State the chosen order (p,d,q), parameter estimates with standard errors and significance, log-likelihood, AIC, BIC, residual SD, the Ljung–Box statistic and p-value, and out-of-sample forecast accuracy (RMSE, MAPE) where applicable. The Section 3 examples on this page provide ready-to-paste templates.

Q10. Can I use this calculator for my thesis or publication?

Yes for exploratory analysis and learning. For the final manuscript, replicate in R (forecast, fable), Python (statsmodels, pmdarima) or Stata, and cite the software and package version. STATS UNLOCK is suitable as an educational reference and as a quick check during model selection.

📚 9. References

The following references support the statistical methods used in this ARIMA model calculator, covering time series forecasting, ACF / PACF interpretation, and best practices in Box–Jenkins methodology.

Box, G. E. P., & Jenkins, G. M. (1970). Time Series Analysis: Forecasting and Control. Holden-Day.
Box, G. E. P., Jenkins, G. M., Reinsel, G. C., & Ljung, G. M. (2015). Time Series Analysis: Forecasting and Control (5th ed.). Wiley. https://doi.org/10.1002/9781118619193
Hyndman, R. J., & Athanasopoulos, G. (2021). Forecasting: Principles and Practice (3rd ed.). OTexts. https://otexts.com/fpp3/
Brockwell, P. J., & Davis, R. A. (2016). Introduction to Time Series and Forecasting (3rd ed.). Springer. https://doi.org/10.1007/978-3-319-29854-2
Hyndman, R. J., & Khandakar, Y. (2008). Automatic time series forecasting: The forecast package for R. Journal of Statistical Software, 27(3), 1–22. https://doi.org/10.18637/jss.v027.i03
Ljung, G. M., & Box, G. E. P. (1978). On a measure of lack of fit in time series models. Biometrika, 65(2), 297–303. https://doi.org/10.1093/biomet/65.2.297
Dickey, D. A., & Fuller, W. A. (1979). Distribution of the estimators for autoregressive time series with a unit root. Journal of the American Statistical Association, 74(366a), 427–431. https://doi.org/10.1080/01621459.1979.10482531
Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., & Shin, Y. (1992). Testing the null hypothesis of stationarity against the alternative of a unit root. Journal of Econometrics, 54(1–3), 159–178. https://doi.org/10.1016/0304-4076(92)90104-Y
Akaike, H. (1974). A new look at the statistical model identification. IEEE Transactions on Automatic Control, 19(6), 716–723. https://doi.org/10.1109/TAC.1974.1100705
Schwarz, G. (1978). Estimating the dimension of a model. Annals of Statistics, 6(2), 461–464. https://doi.org/10.1214/aos/1176344136
Shumway, R. H., & Stoffer, D. S. (2017). Time Series Analysis and Its Applications: With R Examples (4th ed.). Springer. https://doi.org/10.1007/978-3-319-52452-8
Cryer, J. D., & Chan, K. S. (2008). Time Series Analysis With Applications in R (2nd ed.). Springer. https://doi.org/10.1007/978-0-387-75959-3
Hamilton, J. D. (1994). Time Series Analysis. Princeton University Press.
Seabold, S., & Perktold, J. (2010). statsmodels: Econometric and statistical modeling with Python. Proceedings of the 9th Python in Science Conference. https://www.statsmodels.org/
NIST/SEMATECH. (2013). e-Handbook of Statistical Methods — Box–Jenkins models. National Institute of Standards and Technology. https://www.itl.nist.gov/div898/handbook/pmc/section4/pmc44.htm