Strategy ONE

Time Series Forecasting

Tips for Effective Forecasting Questions

  • Be Specific Instead of asking broad questions like "How will our sales perform?", ask about a specific time frame or metric. For example, "What's the projected sales for the next quarter?".

  • Use Natural Language Auto understands conversational language. Frame your questions in a natural way, as if you were asking a colleague.

  • Include Relevant Attributes Include the necessary attributes and metrics in your question to ensure Auto understands the context.

  • Leverage the auto-complete feature For optimal forecasting using Auto, it's recommended to choose metrics and attributes from the auto-complete suggestions. This ensures precise understanding by Auto for accurate forecasting.

  • Avoid Ambiguity Keep your questions clear and unambiguous. Complex or convoluted queries might lead to inaccurate responses.

  • While Strategy enables forecasting of higher-level time units against lower-level data using Auto (such as forecasting next year's cost based on monthly attributes), there are key considerations:

    • Limit on 100 Forecast Points When requesting a forecast of a higher-level time unit against lower-level data, be aware of the 100-point forecast limit. For instance, if your dashboard data is at the daily level and you request a forecast for the next year, the forecast would encompass 1 year's worth of daily data points, exceeding the 100-point limit.

    • Forecast Failure for Incompatible Levels Asking for a forecast of lower-level data against higher-level data results in a forecast failure. For instance, if your dashboard data is at the monthly level and you ask for a forecast for the next week's or next day's values, it won't be feasible due to the data level discrepancy.

Best Practices When Using the Forecasting Analysis Feature

  • Ensure Sufficient Data Volume for Accurate Forecasting

    For more accurate forecasting results, it's important to ensure that your data volume is substantial enough. Behind the scenes, Strategy automatically detects the seasonality of your data. To achieve optimal forecasting outcomes, we recommend that the data you intend to forecast should have at least two complete seasons of historical data.

  • Use High-Quality Continuous Time-Based Data for Forecasting

    While Strategy performs lightweight data processing before forecasting, such as eliminating duplicated data and filling in some missing metric data, it's advisable to conduct forecasting on continuous, high-quality, time-based data. Forecasting results can be compromised if there's a significant amount of missing metric data. It's important to note that forecasts may fail or yield suboptimal outcomes if attributes contain NULL or NaN (Not-a-Number) values.

Optimize Forecast Line Chart Interpretation

Get the most out of your forecast line chart visualization!

  • Hover for Insights Hover your cursor over data points on the forecast line chart to reveal tooltips. These tooltips provide detailed information about forecasted values, as well as upper and lower bounds.

  • Info icon Hover over the info icon to see information about the underlying model and its parameters.

  • Expand for Clarity If you need more detailed information and the ability to display the entire set of data points, expand the visualization. This enhances the readability of the chart.

Troubleshooting

The limitations addressed below are associated with the MicroStrategy ONE Update 11 release. Strategy is actively working on enhancing this feature, and some of these limitations may be addressed in upcoming versions. For the latest information and updates, we encourage you to visit this page periodically. By understanding and considering these limitations, you can make the most of the forecasting feature and achieve meaningful insights.

  • Attribute and Metric Requirements For accurate predictions, make sure to use exactly one attribute and one metric in your questions to be placed in the X and Y axes, respectively. Queries not meeting this requirement may not yield optimal forecasts.

  • Break By Attributes When you include a break by attribute to represent forecasts for multiple elements such as when you forecast sales revenue for the next 12 months for each employee, ensure the number of elements (employees in this case) in the attribute are limited to enhance readability of the analysis results.

  • Consider Attribute Forms The attribute used for forecasting should have at least one form with a type of date, datetime, or integer. Timestamp data is not supported.

  • Timezone-based Attributes Forecasting with timezone-based smart attributes is not currently supported.

  • Consolidation and Grouping Forecasting against consolidated or grouped attribute elements is not possible at this time.

  • Granularity Level Forecasts are not available for less than daily level data (such as hourly, minute, or second intervals).

  • Integer-based Forecasting Be cautious when using integer-based time representations in your queries (such as 202101 for January 2021). Strategy does not automatically convert these time representations to date/time format, leading to unexpected results such as 202113 or 202114.

  • Maximum Forecasting Points You can forecast up to 100 future data points. This limit helps ensure efficiency and accurate predictions.

  • Hindcasting and Tuning Currently, hindcasting against existing data is not supported. Additionally, adjusting algorithm hyper-parameters can only be done via Auto in the current release. For example, you can tune the confidence interval by asking Auto to forecast with a specific confidence level, but this is not available in the dashboard authoring interface yet.

FAQs

What time series forecast method is used in the current release?

The time series forecast method used in our current release is called exponential smoothing. This method supports various non-damped models described on the OTexts site. You can refer to the statsmodel Python module (statsmodels.tsa.exponential_smoothing.ets.ETSModel - statsmodels 0.14.0) for more technical details.

What are the default hyper-parameters used for forecasting?

If not specified by the user, the following default hyper-parameters are used:

  • h value Five data points forward

  • Season model additive (adjustable via Auto)

  • Trend model additive (adjustable via Auto)

  • Seasonality length auto (adjustable via Auto)

  • Confidence interval 95% (adjustable via Auto)

The values for the season and trend models can be set to none, multiplicative, or additive via Auto. You can forecast the next three months of sales with a multiplicative trend model.

Can hyper-parameters be adjusted for forecasting?

Yes, you can adjust hyper-parameters for forecasting through Auto. While these hyper-parameters are not exposed in the Format panel in dashboard authoring, you can fine-tune them using Auto.

What types of attributes and metrics are required for successful forecasting?

To successfully forecast data, you need either:

  • Single attribute of a date or datetime type, plus a single metric of interest (recommended)

  • Single attribute of an integer type, plus a metric of interest (when the date/datetime attribute is unavailable for time series forecasting)

Is missing data handled in forecasting?

Yes, the algorithm can automatically fill in missing data if missing data is less than or equal to 90% of the overall data.

What time-based attribute forecasting levels are supported?

Time-based attribute forecasting is supported for the following levels of data: Daily, Weekly, Monthly, Quarterly, and Yearly.

How does integer forecasting work?

For integer forecasting, the algorithm generates the next integer in the sequence. If an integer is used to represent time (such as 202101, 202102), it is not converted to date/time format, and the forecast adds consecutive integers as future points.