Visitation Forecast is available on our Growth and Enterprise subscriptions. If you're interested in upgrading please reach out to your customer success representative
Contents
Introduction
Our daily visitation forecast provides a 12 month forecast of your daily visitation. Use this information to inform your operational and strategic planning, and allow various departments in your location and third party suppliers (such as cafes and gift shops) to prepare themselves to create the best experience possible for your visitors.
Examples of operational decisions:
Rostering staff such as security, service desk, cafe, gift shop and cleaning staff
Scheduling of certain events or workshops
Purchasing stock levels for cafes and gift shops
Increasing marketing budget or creating special deals on the days when visitation prediction is low
Scheduling maintenance on the days when visitation is low
Scheduling staff training and staff holidays when visitation is low
Examples of strategic decisions:
Budgeting your revenue if you are a paid ticketing location
Providing visibility to your stakeholders and board members of the future
Setting KPIs, Goals, Targets for various departments
How are our forecasts made?
Dexibit's forecasts use machine learning to predict future visitation numbers.
Dexibit takes your location's visitation data from as far back as possible, and combines it with information about events at your location from the Almanac: closed days, the season, the weather and more. We also factor in knowledge about visitation throughout the COVID-19 pandemic.
We then rapidly test many different models against the previous 30 days visitation, and chooses the model with the least error. We use this model to calculate visitation values 12 months into the future.
We produce a new forecast every night so that your visitation predictions are informed by the latest data.
The model error, mentioned above, is provided for your information in the summary section of the Visitation forecast analysis visualization.
Once a forecasted day has passed, the forecasted number will not be updated. Note that because historical forecasts will not change, if historical visitation data is added retrospectively this may cause error to increase.
How accurate are our forecasts?
To assess the accuracy of our forecasts, we provide the error of our forecasts - how far under or over our forecast was compared to the actual numbers. There are a few different ways to calculate error and each is useful in different circumstances. We've explained these below.
Assessing the accuracy of a given day/week/month
Absolute percentage error (APE) is a common way of measuring the accuracy of a forecasted number. It is calculated as follows:
|Actual - Forecast| / Actual * 100 = APE
‘Absolute value’ is notated by "|" in the formula above and means any negative number will be converted to its positive counterpart for calculations. 0% APE means the forecast was very accurate. 30% APE means the forecasted number was 30% either under or over the actual number. It’s unlikely that a forecast would ever be 0%, but by providing sufficient information to Dexibit’s forecast model - such as events in the Almanac - we are able to get a reasonably low APE.
In our visualization you can inspect the absolute percentage error by hovering over each point on the chart. In the expanded view, the APE is given for each row.
Assessing the accuracy of a forecast over a period of time
When it comes to assessing how accurate a forecast was over a period of time, there are three different numbers to look out for:
Mean Absolute Percentage Error (MAPE)
This calculation of error takes the mean average of all APE in a set. This gives you what is known as the mean absolute percentage error or MAPE. MAPE is a commonly used measure of accuracy in demand forecasting, and gives you an idea of how good a forecast is over a period of time. In our product we refer to this as "Mean Error".
1/n * Σ(|Actual - Forecast| / Actual * 100) = MAPE
The lower the "Mean Error" the more accurate the forecast predictions are.
In our visualization the MAPE is given in the summary statistics at the top right of the expanded view.
Weighted Average Percentage Error (WAPE)
One nuance with MAPE is that if your location has quiet days where visitation is low, error in low visitor numbers can have a disproportionate effect on the MAPE and give an inaccurate reflection of the error of a forecast.
If you do have quiet days, weighted average percentage error, or WAPE, gives a better indication of the forecast accuracy over a period of time.
WAPE is calculated as follows:
Σ|Actual - Forecast| / Σ|Actual| * 100 = WAPE
In our visualization the WAPE is given in the summary statistics at the top right of the expanded view.
Model error
Model error provides a holistic measure of the error of the machine learning model in general.
It is calculated as the mean of the absolute percentage errors of weekly totals of the previous 30 days to the forecast running. It excludes days where visitation was less than 30.
As such, it is likely to differ from the MAPE and WAPE of the selected date range. As the model error is provided for the model as a whole, it won't change as the date range changes.
In our visualization the model error is given in the summary statistics at the top right of the expanded view.
How to add a Visitation forecast analysis to your dashboard
Whilst in the dashboards module:
Open the Toolbox on the right hand side
In the Library tab, search for Visitation Forecast Analysis visualization and click Add
Click the edit icon on the visualization to:
- Group the time series by daily/weekly/monthly/quarterly
- change the style of chart
Expanded view explained
Summary panels
The summary panel, at the top right of the expanded view, consists of five metrics. Click on the up and down arrows to cycle through the summaries: visitation; forecast visitation; absolute percentage error; weighted error and model error.
Visitation
Total: the sum of actual visits to your location during the selected time period.
Average: the mean average of visitation for the selected time period.
Max / Min: the highest number of visits and the lowest number of visits for the selected time period.
Depending on your location's business rules, visitation is either calculated through your location's footfall or tickets.
Forecast visitation
Total: the sum of forecasted visitors for your location during the selected time period.
Average: the mean average forecasted visitation for the selected time period.
Max / Min: the highest number of forecasted visitors and the lowest number of forecasted visitors for the selected time period.
Weighted error
Weighted error (WAPE): Weighted average percentage error for the selected time period. The lower the error the more accurate the forecast is.
Total visitation: the sum of actual visits to your location during the selected time period.
Total absolute error: the sum of absolute difference between actual and forecast visitation.
WAPE is explained in more detail above.
Absolute percentage error
Mean error (MAPE): The mean of absolute percentage errors for a given date range. Otherwise known as MAPE. The lower the error the more accurate the forecast is.
High/Low: The highest and lowest absolute percentage errors that were obtained for the selected time period.
MAPE is explained in more detail above.
Model error
Model error: The estimated error of the model chosen for the forecast. The lower the error the more accurate the model is.
Model data start/end date: the date of the first and last record that the model uses.
Model error is explained in more detail above. Note that as the model error is provided for the model as a whole, it won't change as the date range changes.
Table
The table in the expanded view includes the following columns:
Date: The date (or starting date for weekly/monthly/quarterly) of the record
Actual visits: Actual visits on the given date or date range
Dexibit forecast: Forecasted visits in the given date or date range
Absolute percentage error: the absolute percentage error between actual and forecasted (see above)