future – Limitations & Technical Requirements
Last Change: 13/01/2025, 1:30 p.m.
CHECK-IN
- An unlimited number of individual processing runs (e.g., triggered by 'Prepare time series' in step 4 of the frontend) per user account and month is included for normal use. Usage exceeding 500 processing runs is considered abnormal. The data is not permanently provided for user access.
- The more raw data is processed and the more time series are created, the longer the processing takes. If the raw data exceeds 250,000 rows and/or 15 columns or if more than 1,000 time series are created, it may take longer than 15 minutes for processing, or in some cases the calculation may not be executed at all.
- Seconds and time zones in the date column are ignored.
- Only aggregations (e.g., summation) to higher temporal granularities are supported. Disaggregations are not supported, and automatic checks of the smallest temporal intervals in the raw data are not performed.
FORECAST via futureNOW
- The analysis results generated during a session are not stored permanently.
- The number of possible backtesting forecasts depends on the available data history; in some cases, no historical forecasts are possible.
- The quality of the results depends on the underlying data quality, especially the length of the data history. If the data history is very short (less than ten data points), it may not be possible to calculate forecasts at all.
FORECAST via futureEXPERT
- Forecast objects and covariates must be available at the same temporal granularity.
- The forecast period can flexibly be set to a length of up to 60. To achieve a valid result and to calculate backtesting results, the forecast horizon should not exceed a meaningful measure for the given time series. – A rough guideline for a sensible setting: do not set the forecast horizon to more than 1.5 times the typical seasonal length for the given granularity, or max. 1/5 of the length of the data history – whichever comes first.
- If a certain maximum length of the time series is exceeded, only a certain number of past data points are taken into account depending on the granularity of the time series.
- The number of backtesting iterations can be flexibly set to up to 24.
- Forecast intervals are usually calculated for all forecast methods that provide model-based interval boundaries (e.g., ARIMA, TBATS, Exponential Smoothing). In particular, machine learning methods (e.g., SVM, Random Forest, CatBoost) and basic methods such as Moving Average do not output interval boundaries.
- A final ranking and the output of complete results are only provided for forecast methods that were successfully optimized in all backtesting iterations and in the final forecast calculation, and that yield plausible results (automatically identified by plausibility checks). If no successfully optimized and plausible model with complete backtesting results is available, forecast results from a few fallback methods are provided (without backtesting results). This my occur, for example, if the time series is too short for the chosen settings, possibly combined with a high forecast horizon.
- The quality of the results depends on the underlying data quality, especially the length and continuous availability of the data history.
- Replacing missing values is useful for only a few individual missing values. There is no automatic check of the number of missing values. Even if the data history contains too many missing values (in the extreme case, only two values available), these are still replaced. Since this also influences the forecasts, special care should be taken to ensure a good data basis.
- The detection of multiple seasonalities is possible for up to three seasonal lengths. The longer seasonalities must be multiples of the shorter ones.
- The ARIMA forecasting method is available only for forecast objects with monthly, quarterly, or yearly granularity. If explicitly selected for lower granularities, it will automatically be discarded, and a warning will be displayed.
MATCHER
- Forecast objects and covariates must be available at the same temporal granularity.
- An analysis is possible for “smooth” and “erratic” forecast objects, but not for “sporadic” or “lumpy” ones. (More information on the different time series types.)
- A sufficiently long common data history between the forecast object and the covariates is required for the calculation. By default, the overlapping period must be at least four times the typical season length for the respective granularity plus 14 data points. For monthly data, for example, this corresponds to a length of 62 (4*12+14). If the lags to be analyzed are manually set, the minimum length of the common period may be even longer (depending on the maximum lag, possibly in combination with a minimum lag if that is negative). Analyses are only performed for those covariates that meet the conditions for the common time period with the forecast object. All others are automatically excluded.
- Only the results of factors in models with a complete backtesting result can be taken into account (e.g., successful optimization in every backtesting step).
- A bivariate analysis is performed between the forecast object and each covariate individually; multiple influencing factors simultaneously (in combination) or their interactions with each other are not investigated.
- The more missing values the forecast objects or covariates have, the more strongly the analysis results are influenced by the replacement values. Replacement is performed regardless of the number of missing values.
- If a certain maximum length of the time series is exceeded, only a certain number of past data points are taken into account depending on the granularity of the time series.
- The maximum number of temporal shifts (lags) of the covariates relative to the forecast objects that can be tested simultaneously in one run is 15. By default, this is always ensured by the automatic logic. In this process, the lags are tested from -2 up to the minimum of 12 and half of the most typical season length for the granularity (48 for half-hourly, 24 for hourly, 7 for daily, 52 for weekly, 12 for monthly, 4 for quarterly).
POOL
- The factors provided via POOL are not necessarily made available in their original form; instead, they may be altered for analysis purposes (example: disaggregating monthly indicators to daily values). It is not disclosed how the exact indicator is composed.
- There is no guarantee that the factors available in POOL are up-to-date: They are used in the form in which they are available in future at the time of use. This may not coincide with the publication date of new indicator values by the source.
Technical Requirements
- To use the services of prognostica via the SaaS solution future, a device as well as an active internet connection for accessing the software are required.
- futureNOW is optimized for use on a PC with mouse and keyboard. When accessed via smartphones or other mobile devices, some functionality may be limited. To run the software, a compatible browser is required. We recommend and test with Chrome (stable channel), Edge (Stable channel), Firefox (default channel), or Safari in their current versions or the respective two releases prior to the latest version.
- futureEXPERT provides a Python library for communication with the future SaaS platform, requiring a Python environment version 3.9 or higher. (For the provided Jupyter notebooks with examples, Jupyter is required.)
Limitations
- prognostica does not provide a permanent storage system for data and results via future.
- Results from futureNOW are generally provided immediately when using the app or after starting the calculations.
- Results from futureEXPERT are generally retained for one or more days, depending on the type of results.
- The customer is responsible for permanently storing the results generated by futureEXPERT and futureNOW , as well as any data they have uploaded, in their own systems, if permanent storage is desired.
Support Services
- The email address support@future-forecasting.de is available for direct contact with our support team. In particular, this channel can be used to report defects in future and to request additional features in future. In the latter case, prognostica is not obliged to implement and provide these functionalities. Additionally, in the futureEXPERT GitHub project, you can contact us via Issues.
- Other support services, such as training on using the software not included in the directly bookable subscription packages Basic, Standard, and Premium. These must be booked separately if desired, e.g., through the Enterprise package. In addition, if the customer fails to fulfill or does not fully and/or correctly fulfill the required collaboration obligations, an additional fee will be charged for the extra effort incurred. A rate of 108 EUR plus VAT per hour will be charged, with 1/12 of the hourly rate billed for each five-minute unit started.