Login Create Account

Forecast Starter

Time Series β€” Start Here

Your first time series tool. Look at what a series is actually made of, extend it with a rule you can say out loud, then find out whether that rule beats doing nothing at all. No parameters to guess, no Python to load, nothing to install.

πŸ‘¨β€πŸ« Professor Mode: Guided Learning Experience

Never forecasted anything before? Professor Mode walks you through all three steps on a real dataset, one decision at a time.

OVERVIEW & WHAT THIS TOOL SOLVES

Every time-series forecast is a concrete rule for extending a historical line into the futureβ€”yet teams routinely deploy complex forecasting software that fails to beat the simplest possible baseline: "next period will equal last period" (or "this month last year"). Forecast Starter decomposes your series into level, trend, seasonality, and noise, projects six foundational forecasting rules side-by-side, and scores them on a held-out backtest window before you graduate to ARX or ARIMAX.

🎯 Questions It Answers

  • Is our series driven by a steady baseline level, an upward/downward trend, a repeating seasonal cycle, or promotional spikes?
  • Which foundational forecasting rule (Naive, Seasonal Naive, Moving Average, Simple Exponential Smoothing, Drift, or Trend + Seasonality) wins on unseen holdout periods?
  • Does our chosen method actually beat the Naive benchmark (vs. Naive < 1.0), and is it systematically over- or under-forecasting (Bias)?

πŸ“‚ What Data You Need

  • Time-Series Dataset: Ordered date/period column and a continuous outcome metric (e.g., monthly revenue, weekly orders, sign-ups)
  • Cycle & Holdout Settings: Seasonal period length m (e.g., 12 for monthly, 4 for quarterly, 7 for daily) and holdout window size
  • Event Overlay (Optional): Binary promo/event column to inspect whether residual spikes line up with marketing campaigns

πŸ’‘ What You'll Discover

  • Step 1 Decomposition & Event Overlay: Visual breakdown into Trend, Seasonal cycle, and Remainder noise
  • Step 2 Live Forecast & Prediction Fan: Multi-step projection with 80%/95% uncertainty intervals
  • Step 3 6-Method Holdout Leaderboard: Clickable comparison of MAE, MAPE, Ratio vs. Naive, Bias (ME), and RMSE + 1-click handoff to ARX
πŸ” Key Nuances & Common Traps to Watch For
  • Every Forecast Must Beat the Naive Baseline First: If your model's Holdout MAE divided by the Naive method's MAE (vs. Naive) is greater than 1.00, the extra complexity made your forecast worse than simply copying the most recent observation.
  • Never Judge a Forecast on In-Sample Fit: A curve can hug historical data tightly while failing off the right edge of the chart. Always hide the last H periods (Step 3 Holdout Backtest) and score predictions against actuals the model never saw.
  • When to Graduate to ARX / ARIMAX: The six rules on this page use only the outcome series' own history. When Remainder spikes line up with promotions, price changes, or ad bursts, click Open in ARX to bring in external predictors and autoregressive lags.
πŸ“ Under the Hood: Model Equations & Variable Definitions

1. Baseline & Smoothing Forecast Rules (Naive, Seasonal Naive, Moving Average, SES)

\[ \hat{y}_{T+h}^{\text{Naive}} = y_T, \qquad \hat{y}_{T+h}^{\text{SNaive}} = y_{T+h-m\,\lceil h/m \rceil}, \qquad \hat{y}_{T+h}^{\text{MA}(k)} = \frac{1}{k}\sum_{i=0}^{k-1} y_{T-i}, \qquad \ell_t = \alpha y_t + (1 - \alpha)\ell_{t-1} \]

πŸ’¬ Read it in plain English: Naive repeats the final observation y_T, Seasonal Naive repeats the same season from m periods ago, Moving Average averages the last k periods equally, and Simple Exponential Smoothing (SES) updates level β„“_t by blending the latest actual y_t (weight Ξ±) with the previous level (weight 1 βˆ’ Ξ±).

yt, T, h
Historical Series, Origin & HorizonYOUR DATA β“˜
Observed value at period t, last training period T, and forecast step ahead h = 1, …, H.
m, k
Seasonal Cycle & MA WindowYOUR DATA β“˜
Number of periods per seasonal cycle m (e.g., 12 or 4) and moving-average window length k.
Ξ±, β„“t
SES Smoothing Weight & LevelESTIMATED β“˜
Smoothing parameter Ξ± ∈ (0, 1) and smoothed local level β„“_t.
Ε·T+h
h-Step-Ahead Point ForecastOUTPUT β“˜
Projected value h periods after the training cutoff T.

2. Trend + Seasonality Decomposition & Holdout Backtest Metrics

\[ \hat{y}_{T+h}^{\text{TS}} = \left(\hat{\beta}_0 + \hat{\beta}_1 (T + h)\right) + \hat{s}_{(T+h) \bmod m}, \qquad \text{MAE} = \frac{1}{H}\sum_{h=1}^{H} |y_{T+h} - \hat{y}_{T+h}|, \qquad \text{vs. Naive} = \frac{\text{MAE}_{\text{method}}}{\text{MAE}_{\text{Naive}}} \]

πŸ’¬ Read it in plain English: Trend + Seasonality extends a linear time trend Ξ²Μ‚β‚€ + β̂₁t and adds the repeating seasonal index ŝ for that month/quarter, while the Holdout Backtest scores accuracy over H hidden periods relative to the Naive benchmark.

Ξ²Μ‚0, Ξ²Μ‚1, ŝm
Trend Intercept, Slope & Seasonal IndexESTIMATED β“˜
Fitted baseline level, per-period trend slope, and zero-sum seasonal offset for season m.
MAE, RMSE, MAPE
Holdout Forecast Error MetricsINFERENCE β“˜
Mean Absolute Error, Root Mean Squared Error, and Mean Absolute Percentage Error over H holdout steps.
vs. Naive
Relative Error Ratio (Theil’s U₁ analog)INFERENCE β“˜
Method MAE divided by Naive MAE (< 1.00 beats the Naive benchmark; > 1.00 loses to Naive).

DATA SOURCE

πŸ“š

Use a Case Study

Four series, each one teaching a different answer to "which method should I use". Start with the first β€” it is the one where the cleverest method loses.

πŸ“€

Upload Your Data

A CSV with one column of dates or period labels (ordered chronologically) and one continuous numeric column (interval/ratio scale: revenue, units, subscribers, CPC), with no missing periods in the middle.

Drag & drop a CSV (.csv, .tsv, .txt)

First row = column headers. One row per period, oldest first (50+ periods recommended for seasonal cycles).

Need help formatting your CSV? Open Import Wizard β†’