Using DClaw Forecast.

1. Bring your data

In the Workspace, open Data and drop a .csv or paste a date and a value on each line. Monthly, weekly and daily data all work: forecasts here are monthly, so weekly and daily values are combined into calendar months, as a total (sales, visits) or an average(prices, temperature). Partial months at either end are left out; a gap in the middle is reported so you can fill it. Daily data with no weekends (trading or working days) is judged against each month's weekdays.

The preview says how the file was read before anything is loaded: the spacing it found, how dates like 03/04/2024 were read (day first or month first, decided once for the whole file), which column the values come from (pick another if the file has several), and any repeated dates. Semicolon-separated files may use a comma as the decimal point. A plain list of numbers, one per line, also works; it is read as months ending last month.

A series needs 8 to 600 months. Imported series are saved to your account (up to 50); files up to 5 MB.

2. Read the forecast

The solid line is your history and the dashed line is the forecast. The shaded band is the range at the confidence level you chose (80, 90 or 95%): the forecast expects the real value to land inside it that often. It widens further out because uncertainty grows with the horizon. Set a target and the forecast gives the chance of reaching it at the end.

3. Models

• Linear: a straight-line trend plus the average shape of each month.
• ETS (Holt-Winters): level, trend and season that adapt as new months arrive. Needs two full years; with less it is fitted as Linear, and says so.
• Theta: smoothing plus half the long-run drift, on the seasonally adjusted series.
• Seasonal naive: each month repeats the same month last year. The benchmark to beat.
• ARIMA: seasonal ARIMA with the order chosen automatically.
• Ensemble: Linear and ETS blended, weighted by their recent accuracy.
• Auto: whichever of the six had the lowest error on the most recent 3 to 6 months, re-fitted without them (a single holdout).

Prophet, Random Forest and XGBoost are planned, not available.

4. Accuracy measures

Every score is measured on months the model did not see: it is re-fitted without them and its predictions are compared with what happened.

• MAPE: average miss as a percentage of the actual value. Months whose actual value is 0 have no percentage and are left out.
• SMAPE: a percentage that also works near zero.
• MAE and RMSE: average miss in the series' own units; RMSE weighs big misses more.
• Coverage: how often the real value landed inside the range, against the level it claims.
• Direction: how often the forecast called the month-to-month move up or down correctly.

The Lab's Backtest view repeats this from several starting points (rolling-origin cross-validation). With too little history to hold months back, the score reads “not enough history to validate”.

5. Settings you can change

Model, horizon, confidence level, the ETS smoothing weights, how much history to learn from (all, or the last 30 or 24 months), and scenario assumptions: a trend tilt, seasonal strength, a one-off shock, and a timed change with its own size, start, length and shape (step, gradual, temporary or custom), returning to normal or staying afterwards. Percentage changes act on the size of a value, so +10% raises a negative forecast too. Every setting explains itself in the app; expected effects are guidance, observed effects are always calculated.

6. The Forecast Lab

The Lab (/app/lab) is for changing one thing at a time against a baseline: Experiment, Guided experiments, Sensitivity sweeps, two-setting Heatmaps, Relationships, Scenarios, Backtest, Models, Importance, the real-data Domains and your History (saved to your account, newest 30).

• Beat the model: six real months are hidden; draw your guess and a range, and see whether you or the model came closer. Each set of hidden months counts once toward your record.
• Time machine: re-fit on only what was known at a past month and replay the forecast against what happened.
• Say it in words: type a what-if (“+10% for 3 months from month 4”); it becomes settings you can see and edit.
• 1,000 possible futures: simulated paths that match the range, with a live chance of reaching a target.

7. Real data: weather, GDP, markets

Weather: observed monthly weather for five Indian cities from the Open-Meteo archive (ERA5 reanalysis, CC BY 4.0): high temperature, rainfall, cloud cover, peak wind, humidity and sunshine, each forecast as a low, middle and high case. GDP: official US quarterly statistics from FRED (real GDP, the Fed funds rate, CPI inflation, unemployment and the Philadelphia Fed manufacturing survey). Markets: daily closes of the S&P 500, Nasdaq Composite and Dow from FRED. Each keeps a labelled sample you can switch to. In the Workspace, Real data offers New Delhi temperature and the S&P 500 monthly average.

8. The Copilot

Ask the Copilot about the forecast in front of you, in the Workspace or the Lab. It answers only from that forecast's numbers and says when Claude answered. Each account gets 30 AI answers a day; after that, and whenever no AI service answers, a built-in explainer replies instead (it doesn't use up an answer). AI answers can be wrong: check them against the numbers on screen.

9. Your data and account

The account menu's Your data & account page lists what is stored, downloads all of it as one JSON file, and deletes your account and everything in it. Limits per account: 50 saved series, 100 saved forecasts, the newest 30 Lab experiments and 500 Beat-the-model rounds. Your data is never visible to other users; see Security and Privacy.