Forecast revenue, spend, orders, conversions, or demand from your historical time series. Choose the cadence, horizon, and level of detail. Get a forward view with the assumptions and caveats still attached.
History through today
30-day horizon
Define the metric, the history it should learn from, the cadence you operate on, and how far forward you need to see. The workflow preserves those choices instead of quietly changing them.
What future value are we estimating?
Enough signal to expose trend and seasonality.
Hourly, daily, weekly, monthly, or yearly.
Future periods measured in that same cadence.
One company line or many independent series.
Build base, upside, and downside views from explicit planning assumptions. Compare them on the same horizon and frequency. Never present scenarios as one objective truth.
Validate the timestamp, target columns, frequency, date range, and number of series before the model runs.
| Date | Series | Revenue |
|---|---|---|
| 2026-05-01 | Store | $42,814 |
| 2026-05-02 | Store | $39,672 |
| 2026-05-03 | Store | $46,105 |
Forecast one or more metrics into the future, with the horizon measured in the prepared data’s hourly, daily, weekly, monthly, or yearly frequency.
Use the output as evidence for the business decision, then state the assumptions, horizon, and confidence level next to the recommendation.
Use the median path for normal operating targets.
Identify the inventory or cash threshold that matters.
| Date | Series | Revenue | Promo % |
|---|---|---|---|
| 2026-05-01 | Store | $42,814 | 9.2% |
| 2026-05-02 | Store | $39,672 | 7.8% |
| 2026-05-03 | Store | $46,105 | 12.1% |
| 2026-05-04 | Store | $44,388 | 8.7% |
| 2026-05-05 | Store | $48,210 | 10.4% |
| 2026-05-06 | Store | $45,941 | 9.0% |
Preparation catches missing dates, mixed grain, wrong columns, and sparse series before they become a misleading forecast.
Use the central path as the normal operating plan.
Define the inventory, cash, or capacity threshold that matters.
Decide what extra demand would trigger more spend, stock, or staffing.
The forecast informs the decision. It does not make the operational change by itself.
Moby can use Linear, Seasonal, Prophet, and Chronos, or select the algorithm automatically based on the data and forecast requirements. Linear, Seasonal, and Prophet are classical forecasting approaches. Chronos is a transformer-based time-series foundation model: the same broad architecture family that powers GPT, applied to sequences of numeric observations instead of sequences of words.
Triple Whale used more than $100 billion in ecommerce GMV to fine-tune and calibrate the forecasting system for how commerce actually behaves: promotions, launches, retention curves, holiday spikes, inventory constraints, channel shifts, and changing growth stages.
The result is not an off-the-shelf model fitted to one store in isolation. It combines the selected forecasting approach with Triple Whale’s proprietary ecommerce training context.
A clear baseline for consistent directional trends where the signal does not need a more complex seasonal model.
Consistent trendDesigned for time series with repeating cycles, recurring peaks, and a cadence that reliably returns.
Repeating patternUseful when the series carries multiple seasonalities, holidays, changes in trend, or missing observations.
Seasonality + gapsA transformer-based time-series foundation model that tokenizes numeric history, attends across the sequence, and predicts probabilistic future values instead of the next word.
Time-series transformerBoth use transformer architecture to learn relationships across a sequence and predict what comes next. GPT works with language tokens. Chronos converts time-series values into tokens and predicts a distribution of future values, which is why it can transfer patterns learned across many different series.
Words become tokens. Attention connects the context. The model predicts the next token.
Numeric observations become tokens. Attention connects the history. The model predicts future value ranges.
Moby can route the request to the forecasting approach that best matches the history, cadence, missingness, seasonality, and horizon.
Base, upside, and downside scenarios stay on the same cadence and horizon so the comparison is useful. Each path keeps its assumption in plain view.
Illustrative index only. This is not shop data or a guaranteed outcome.
Confidence comes from the quality of the history, the regularity of the series, and a horizon that is reasonable for the signal available. The honest answer is sometimes directional. Sometimes it is low confidence. That distinction matters.
Longer, relevant history gives trend and seasonality more room to show themselves.
Missing timestamps, mixed grain, and sparse series weaken the planning signal.
A moderate horizon is more defensible than asking a short history to predict too far ahead.
Launches, pricing shifts, channel changes, and shocks can break the pattern the model learned.
Change the metric, cadence, horizon, or series population without losing the history the forecast learned from or the assumptions attached to the result.
Use the output to create a planning brief, set targets, size inventory, stage spend, prepare staffing, or explain the range to leadership.
Forecasting is decision support. It does not automatically purchase inventory, change budgets, or execute campaigns.
Revenue, spend, orders, conversions, product demand, or another defensible numeric time series. The exact metric, grain, date window, and series population should be defined before the model runs.
There is no universal minimum that guarantees quality. Confidence improves when there is enough relevant history for trend and seasonality to repeat. Sparse history, mixed cadence, and structural changes reduce reliability.
The forecasting workflow supports one to 365 future periods, measured in the prepared frequency. A 30-period daily forecast means 30 days; a 12-period monthly forecast means 12 months. Longer horizons should be interpreted more cautiously.
Yes, when each item is represented as a usable independent series with enough history. Multi-series forecasting can repeat the same workflow across series IDs, and sparse series can be included or excluded explicitly.
Historically aligned covariates can be included when they are available and defensible. Avoid leakage, and do not treat unknown future covariate values as if they are already known.
No. Time-series forecasting projects future values from historical patterns. MMM estimates the incremental relationship between marketing inputs and a KPI and can simulate budget allocation. They answer different questions and should not be presented as interchangeable.
Define the metric, horizon, cadence, and decision. We will show you how forecasting turns the history into a forward view without hiding the assumptions.