future - Services and Pricing

• Handling of raw data in csv or tsv format with UTF-8 encoding, via futureEXPERT: additionally pandas dataframe.
• Possibility to define value and grouping variables/columns as well as a date variable/column. (At least one value and exactly one date column must be defined, grouping and additional value columns are optional.)
• Support of numeric date formats, e.g., YYYY-MM-DD, DD.MM.YYYY hh:mm:ss.

• Checking data plausibility concerning defined data types, missing values, and set thresholds.
• User-defined response to plausibility violations per column and type (e.g., deletion or setting replacement values). Convenient configuration only in the frontend, with the downloaded configuration file allowing seamless continuation in the Python client.

• Option to use own factors on a daily or monthly basis.
• Option to use factors from POOL (see below).
• Manual selection of different lags per factor (up to 30 days or 12 months into the future or past).

• Manual definition of lags per covariate and time series or use of MATCHER results for lag selection, with manual adjustment options.
• Exponential smoothing as forecast method for covariates
• All covariates in a joint model

• Option to choose an additional method for incorporating factors into the forecast model.
• Choice of whether factors should be considered individually (single) or jointly (joint) in the forecast model.

• Default: typical season lengths for the given granularity are checked.
• Alternatively, specific season lengths (including multiples) can be specified for checks.
• A fixed season length can also be set, by passing the season detection..

• Level shifts
• Significant structure changes in the occurrence of very small values or zero values

Calculation of rolling historical forecasts for all suitable forecasting methods, aligned with the set forecast horizon, with preset values suited to the time series (length, granularity). Options include:

• Defining the number of iterations
• Shift length
• Refitting on each iteration (significant impact on runtime)

• Preselection based on preprocessing-identified time series characteristics
• Final ranking based on backtesting results
• Option to set one accuracy measure for each of the two time series groups (smooth/erratic and sporadic/lumpy), which is used as ranking and selection criteria.
• Forecast steps with weighting (step_weights) for ranking can be set.

• Time series with up to one covariate
• Forecasts including data history, possibly with prediction interval (per successful model) and certain preprocessing results
• Backtesting forecasts (one plot per model and iteration)

• Overview of certain preprocessing results and information about the best model per time series (as data frame)
• Forecasts, possibly with prediction intervals, per best model and time series (as data frame)

relating to publication delay and the time offset between forecast object and factors:

• Option to exclude covariates with no temporal lead relative to the last data point of the forecast object (and thus unsuitable for any forecast step).
• Option to set the maximum allowable time difference between the most recent value of the influencing factor and the forecast object.

• Ranking of successful factors with best lag per forecast object including benchmark model.
• Time series values of forecast objects.
• Data history of ranked covariates.

for use in other modules. Examples include:

• Various daily weather data at selected german weather stations
• Holiday & workday information
• Selected economic indicators

• Working days as in Würzburg, Bavaria (Germany)
• Indicated times are based on the current time in Berlin, Germany