Forecasting, explained.
Don't just see the forecast. Understand what moves it.
DClaw Forecast studies historical data to estimate what may happen next — then lets you change assumptions and see how the prediction responds.
Use it for economic indicators, markets, weather, sales, demand and other time-series data.
A small working lab. Change one thing and watch the forecast respond. Or watch a two-minute silent tour of the whole app.
Scenario active for 4 quarters. After the period: return to baseline.
Calculated from this demo run. Change anything to run a new experiment.
Example data, not live economic statistics. Demo rule: each +1 pp in the rate lowers growth by up to 0.35 pp; the effect builds over the first quarters of the window and halves every quarter after it ends. Invented for illustration — not an estimated economic relationship.
This is the kind of question DClaw Forecast is built for. Here is what it does with it, using the demo above.
Demo figures, not real economic data. *The past-data score is the Lab's GDP model tested on 8 held-out quarters of its sample data, against assuming average growth.
Six steps, the same for every kind of data.
Start from a built-in example or upload your own history as a CSV.
Monthly, weekly or daily data. Weekly and daily values are combined into months, because forecasts here are monthly.
Pick the number you care about.
The model estimates what may happen next if nothing in your experiment changes.
Change an assumption, a model setting or how far ahead to look, plus when a change starts and how long it lasts.
See the baseline and your experiment side by side.
What changed and when, how uncertain it is, how well the method did on past data, and why it responded.
A forecast is an estimate of what may happen in the future, based on patterns in historical data and, where supported, additional inputs or assumptions.
A forecast is not a guarantee. That's why DClaw also shows:
Finds patterns such as:
The expected future path
The range of plausible outcomes
The steps stay the same, but the experience adapts to the kind of data: quarterly economics, daily markets, uncertain weather, seasonal business numbers.
In the LabWith your own monthly dataNot yet available
“What if interest rates stay higher for longer?”
Quarterly data. In the Lab it runs on official US statistics. You set the future path of interest rates, inflation, business activity and unemployment, then see which input moved GDP and by how much.
“Can yesterday's move predict tomorrow's?”
Daily prices (the S&P 500, Nasdaq and Dow in the Lab), technical features like moving averages and RSI, and every model scored on days it never saw, so you can tell a real signal from noise.
“How much less certain is the forecast a year out than next month?”
Six measures (high temperature, rainfall, cloud cover, peak wind, humidity and sunshine), each as a range instead of a single number (low, middle and high cases), with each month's sky and season read from the forecast. In the Lab it runs on real observed weather for five Indian cities.
“What if demand jumps during the holiday period?”
Monthly data with a yearly rhythm. Add a promotion or a one-off event, choose when it starts and how long it lasts, and compare.
“How much demand might we need to plan for next month?”
The same monthly engine read as a planning range: the middle is the likely need, and the shaded range tells you how much buffer to hold.
Two kinds of change, kept visibly apart.
Change the scenario you want to test.
Change how the forecasting method learns from the data.
forecast horizontrend smoothing (β)season smoothing (γ)modeltraining windowThe same change can produce a different model response depending on when it happens and how long it lasts.
Three demo experiments on the GDP example, each starting 2027 Q1. In the Lab you choose the start, the length, the shape and what happens afterwards.
The shaded area shows a range of outcomes the model considers plausible. It gets wider the further ahead you look, because small errors add up over time.
Technical: prediction interval, or a quantile range such as P10–P90.
Every change is listed, and so is everything that stayed the same. The response is calculated, never guessed.
Demo run. The difference is calculated; no cause is claimed beyond the assumption you changed.
Before running it · guidance only
After running it · calculated
Before trusting a forecast, test it. Hide the most recent months, let the model predict them, compare with what actually happened, then move forward and repeat. Below is the Lab's real test on the Electronics Demand sample data (ARIMA model).
Technical name for this kind of test: rolling-origin backtesting (time-series cross-validation).
Hover the info icon for what a setting is, why it matters and what happens if you raise or lower it. Or switch on Help mode and every term explains itself where you meet it.
ets.betaMove the slider to see what the change is expected to do — the Lab then measures what it actually did.
Real parameter from the Forecast Lab's registry · ETS · Ensemble
Prediction error that penalises larger misses more strongly.
How often the actual value fell inside the model's interval.
The shaded range showing where future values may reasonably fall.
How quickly ETS updates its estimate of the growth rate.
With Help mode on, parameters, metrics and chart layers explain themselves in place — no leaving the app to look up a term. Values shown are the Lab's Electronics Demand sample.
You don't need to be a data scientist. Start in Beginner mode; switch to Advanced when you want more control.
Slide one setting across its range and re-run the forecast at every value. Computed from the Electronics Demand sample data with the ETS model. Technical name: sensitivity analysis.
Different methods make different trade-offs. Compare them on the same historical test. Scores below are real, from the Electronics Demand sample data.
| Method | Average errorAverage errorAverage size of the prediction error.Technical name: MAE | Large-error sensitivityLarge-error sensitivityPrediction error that penalises larger misses more strongly.Technical name: RMSE | Range coverageRange coverageHow often the actual value fell inside the model's interval.Technical name: coverage | Status |
|---|---|---|---|---|
| Linear | 4,457 | 4,630 | 3/12 | In the Lab |
| ETS (Holt-Winters) | 3,536 | 3,705 | 10/12 | In the Lab |
| Theta | 4,055 | 4,461 | 5/12 | In the Lab |
| Seasonal naive | 21,582 | 21,599 | 12/12 | In the Lab |
| ARIMA | 8,249 | 8,854 | 8/12 | In the Lab |
| Ensemble | 3,996 | 4,159 | 8/12 | In the Lab |
| Prophet | — | — | — | Planned |
| Random Forest | — | — | — | Planned |
| XGBoost | — | — | — | Planned |
Most forecasting tools show you a line. DClaw helps you answer:
Try them in the Forecast Lab (free, sign in with email or Google).
The goal isn't to make the future certain.
It's to make the forecast understandable.
Build a baseline, change an assumption and understand how the model responds.
Economy, markets and weather are in the Lab, plus business and operations data on monthly samples or your own uploads.