Top 30 time-series forecasting models & methods

Forecasting looks magical when it works: you show a line chart, draw the future, and everyone nods. But the real skill is not “which algorithm is best?” The real skill is knowing what kind of time series you have—and picking the simplest method that matches it.

A time series is just data that arrives in order: daily sales, weekly signups, hourly demand, monthly revenue, call volume per minute, temperature, stock levels, churn events. Forecasting means predicting the next values based on patterns like trend, seasonality, cycles, and random noise—plus real-world shocks like campaigns, holidays, outages, and pricing changes.

This blog is a student-friendly list of 30 forecasting models and methods (from classic to modern), with guidance on when each one makes sense.

Before models: the 30-second checklist

Most forecasting mistakes happen before modeling. So quickly check:

1) Frequency & horizon
Are you forecasting hourly, daily, weekly, monthly? And how far ahead—next day, next month, next quarter?

2) Trend
Does the baseline rise/fall over time?

3) Seasonality
Does it repeat weekly/monthly/yearly? (e.g., weekends, month-end, festive seasons)

4) Multiple seasonalities
Some series have more than one seasonality: hourly + weekly, or weekly + yearly.

5) External drivers (events)
Price change, ads, holidays, product launch, policy change, supply constraints—these often matter more than the algorithm.

6) Data length
Many models need enough history (especially seasonal models). If you have very little data, simpler methods win.

Forecasting sounds complicated, but at its core it’s just answering one simple question: what’s likely to happen next, based on what’s happened before? Businesses use forecasting every day, predicting next week’s sales, next month’s revenue, tomorrow’s website traffic, call center volume, inventory needs, or even how many leads a campaign will generate. The “models” below are just different ways to make that guess more reliable. Some are super simple (and surprisingly strong as baselines), some handle seasonality and trends better, and some use extra signals like promotions, holidays, price changes, or weather. You don’t need to be a data scientist to understand them, use this list as a menu: pick a method that matches your data pattern and your business reality.

1. Naïve Forecast — Assume the next value will be the same as the latest value. Example: tomorrow’s website visits = today’s visits for a stable site.
2. Seasonal Naïve — Assume the next value will match the same day/week/month from the last cycle. Example: next Monday sales = last Monday sales.
3. Moving Average — Use the average of the last few days/weeks to guess the next value. Example: forecast daily orders using the last 7-day average.
4. Weighted Moving Average — Same as moving average, but recent days matter more. Example: forecast leads where the last 2 days matter more than last week.
5. Simple Exponential Smoothing (SES) — A smart moving average that automatically gives more weight to recent data. Example: forecast support tickets/day when there’s no big seasonal pattern.
6. Drift Method — Assume the trend continues at the same pace as before. Example: forecast monthly subscribers if you’re adding ~500/month consistently.
7. Holt’s Linear Trend — Like drift, but it learns the trend more smoothly. Example: forecast growing app signups during a steady expansion phase.
8. Holt–Winters — Handles both trend and seasonality together. Example: forecast weekly sales that rise over months and spike on weekends.
9. ETS family — An automated version of smoothing that picks the best trend/seasonality style for your data. Example: forecasting monthly revenue with regular seasonality.
10. AR (Autoregression) — Uses recent past values to predict the next one (yesterday affects today). Example: electricity usage or daily demand where the last few days matter.
11. MA term (in ARIMA) — Learns from past “mistakes” to correct future forecasts. Example: demand forecasts that often get thrown off by one-time spikes.
12. ARMA — Combines “recent values” + “past errors” for stable series. Example: stable signals like system latency without strong seasonality.
13. ARIMA — Good for data that drifts up/down over time but isn’t strongly seasonal. Example: forecasting inventory levels or cashflow with a trend.
14. SARIMA — ARIMA plus seasonality (weekly/monthly/yearly repeating patterns). Example: monthly bookings that repeat every year or weekly traffic cycles.
15. ARIMAX / SARIMAX — SARIMA plus outside factors (like price, campaigns, holidays, weather). Example: forecasting sales with promotions + holidays included.
16. Theta Method — A simple, reliable method that often works great for trending business data. Example: forecasting monthly sales for many SKUs as a strong baseline.
17. STL Decomposition + Forecast — Split the data into trend + seasonality, forecast each, then combine. Example: website traffic with clear weekly seasonality + long-term growth.
18. State Space / Kalman Filter — Continuously updates the forecast as new data arrives (real-time adjustment). Example: delivery ETA tracking or sensor-based monitoring.
19. Linear Regression with time features — Use calendar patterns (weekday, month, holiday) + history to forecast. Example: forecast leads using day-of-week and campaign calendar.
20. Dynamic Regression — Regression + a “time correction” so patterns don’t get missed. Example: forecast sales from ad spend, while also handling natural weekly cycles.
21. VAR — Forecast multiple related things together because they influence each other. Example: product A and product B sales move together, so predict them jointly.
22. VARMAX — VAR plus outside drivers. Example: forecast region-wise demand + marketing spend together.
23. Cointegration (VECM) — Used when two things move together in the long run even if they fluctuate short-term. Example: currency pairs, oil price and shipping cost.
24. Granger Causality (test) — Checks if one thing helps predict another (a “does it help?” test). Example: does ad spend help predict signups better than signups history alone?
25. Random Forest (time-series features) — A smart “rule-based” predictor that handles complex patterns using many inputs. Example: forecast delivery time using traffic, weather, and past delays.
26. Gradient Boosting (XGBoost/LightGBM/CatBoost) — Often the best practical business forecaster when you have lots of factors. Example: store-level demand using price, promos, weather, stockouts.
27. Prophet — Built for business time series with trend + seasonality + holidays. Example: forecast marketing leads with holiday spikes and seasonal dips.
28. TBATS — Great when you have multiple repeating patterns at once. Example: hourly demand with daily and weekly cycles together.
29. LSTM / RNN — Deep learning that learns patterns in long sequences when you have a lot of data. Example: IoT sensor streams or very large demand datasets.
30. Probabilistic Forecasting — Gives a range, not one number (best case / likely / worst case). Example: inventory planning: “We’ll likely sell 1,000 units, but it could be 800–1,300.”

Which model should you start with?

If you’re a student or building your first forecasting project, start like this:

• If your series is stable → Seasonal Naïve, then ETS/Holt-Winters.
• If trend + seasonality is strong → ETS or SARIMA.
• If promotions/holidays matter → SARIMAX or Boosting (XGBoost/LightGBM) with event features.
• If multiple seasonalities (hourly + weekly) → TBATS or decomposition + ML.
• If you have many related series → VAR or ML with shared features, and consider hierarchical forecasting.
• If decision-making needs confidence → add prediction intervals (probabilistic forecasting).

---

The 5 mistakes that break most forecasts

1) Not using a time-based split
Don’t randomly split time series. Train on past, test on future.

2) Ignoring seasonality
If there’s weekly seasonality and you ignore it, your model will repeatedly “surprise” itself every weekend.

3) Forgetting events and business changes
Marketing campaigns, pricing, stockouts, policy changes—these create breaks. Your model can’t read your business calendar unless you feed it.

4) Forecasting noisy series too far ahead
Some series are inherently unpredictable beyond a short horizon. In that case, focus on ranges and scenarios.

5) Measuring with the wrong metric
MAPE can be misleading when values are near zero. Sometimes RMSE, MAE, SMAPE, or pinball loss (for quantiles) makes more sense.

How SCIKIQ fits

Forecasting isn’t only about picking ARIMA vs Prophet vs LSTM. In real companies, forecasting fails for a simpler reason: teams don’t agree on what the numbers mean. “Revenue,” “demand,” “orders,” “active users,” “stock available”—the definitions change across systems and departments, and forecasts become untrusted.

That’s where SCIKIQ helps. SCIKIQ builds a governed layer of business context, definitions, KPI logic, metadata, and relationships, so your forecasting inputs are consistent and your outputs are explainable. When your data meaning is anchored, models become easier to deploy, forecasts become easier to trust, and planning becomes faster across the organization.

Also read. https://scikiq.com/blog/how-statistical-models-help-you-do-better-sales-predictions/