How to operationalize forecasts with machine learning
- In Blog, Leadership
- ai, forecast, goals, machine learning, ML, planning, plans, targets
- 5 min read
In visitor attractions, forecasting using machine learning makes a huge difference to strategic and operational planning. With AI, these predictions can be made faster, with less work, while improving accuracy. But getting forecasts and getting your team using forecasts are two different challenges.
Here, we’ll talk through what you need to operationalize forecasts in strategic and operational planning processes.
How it works
Before we get into how to use machine learning: a quick primer on how it all works. Machine learning models are a branch of Artificial Intelligence (AI). In Dexibit, we have a series of models to predict different metrics (such as visitation versus revenue versus attrition) in different scenarios (such as experience versus event based venues) in different ways (such as short term versus long term). We use these models globally across our customers, trained on the customer’s own data, whereby the model adapts its features specifically to that location based on what’s evident in the data.
These models are fed historic data for the metric, third party data (such as weather), contextual data (such as the almanac of what’s on in and around the location) and Dexibt’s global unified data of what this looks like across industry. Apart from the metric we’re predicting, we should usually have both historic and future data in order for it to be predictive. For example, if we’re using ‘free days’ as a feature (this is what a factor the model uses is called), we need to have marked in the almanac days in the past where we had a free day, then days in the future where we’ve got another scheduled, in order for the model to determine the impact of free days and then make a prediction for the next.
Alongside predictions, these models output feature importance analysis, which is useful insight into why they’ve made a prediction, plus their accuracy. This helps with transparency, which in turn helps trust – an important part of your team’s adoption of forecasting so they can be confident in the numbers.
Forecasts don’t replace your goals, targets or plans – they’re a tool to help set and evaluate them. Forecasts are what is likely to happen, goals are what we want to happen. Both should coexist side by side – forecasts will update every day, whereas targets will (hopefully) stay static across a reporting period such as a year or a quarter.
This leads way to the 5 moments to operationalize forecasts in visitor attractions:
Strategic planning
Typically, when a visitor attraction is approaching their planning season in the lead up to the new financial year, finance will lead an effort in conjunction with other stakeholders throughout the organization to set the year’s financial plan. For the top line, these projections are based off an assumption of how many visitors will come and when.
Traditionally, this might be manually calculated based from deep analysis to compare previous years within the context of what was happening at the time that may have influenced these numbers. For example, a public holiday might be on a Monday one year, then a Wednesday in another, delivering quite different results.
Using machine learning, this base work can be automated and the input to the strategic plan considered alongside this context to evaluate it. The team will still have the opportunity to dial up or down their level of confidence in the numbers to either add stretch or contingency – or both, in the form of a modest financial plan (potentially monthly granularity) and then a more challenging target set (potentially daily or weekly granularity). Automating the first pass of this plan can usually save around a week’s analysis.
Reprojections
Often, things change through the year, such as marketing budgets or event schedules. This can sometimes prompt a team to consider reprojecting their financial plan for the coming quarter if it is considered too ambitious or conservative as the year has unfolded.
Because machine learning models update automatically every day, they’re ready to go whenever the team need to reproject, working in a similar way to when targets are initially set. Additionally, by comparing the target or plan together with actual performance to date for the period and the forecast for the future time remaining, it’s possible to track and prompt when reprojection may need to occur. Reprojections usually save around a day’s work each time, so nearly another week’s work over the course of the year.
Operational planning
As the year unfolds, public program and front line teams will need to perform operational planning. This is usually on two time periods – around 90 days in order to scale up or down labor forces for the season, and around 14 – 21 days in order to inform workforce rostering and inventory scheduling.
Machine learning models keep updating every day and will usually become more accurate the closer the prediction is to the day of performance, especially with the help of more certain data such as event schedules and new data such as weather forecasts (that are only available a week or two out). As a result, the operations team will benefit from the most accurate forecasts right at the time they need them most. Sometimes, the visitor attraction might also share these forecasts with third party operators such as a food and beverage partner running the cafe.
Because operational planning happens so often, usually being tweaked every week, operationalizing forecasts in this way can save a significant amount of time, which usually adds up to several week’s work over the course of the year.
Dynamic pricing and capacity control
For attractions using variable or dynamic pricing structures, or controlling capacity with overage allowances, forecasting can be used towards these controls, either manually or programmatically (via outbound integration). Either way, a watchful eye is important to ensure pricing guardrails are applying as intended.
This achieves a cost or time saving over manual analysis to support dynamic pricing and provides more accurate predictions in order to support revenue optimization.
Review and attribution
Once the performance period of a forecast has passed, forecasts continue to be useful. Users may enjoy tracking the forecasts actual performance, which helps build trust in the model or identify areas for improvement so the model can be further refined. A common tweak which aids forecast accuracy is to capture almanac context which may have been missed, such as a new event schedule.
Between what the forecast predicted and what actually happened is the residual value. These values can aid attribution, the challenge of working out whether an increase or decline in visitation was correlated to something the visitor attraction did differently, or a change in the world around it. For example, if visitation exceeds forecast for a day on which a new offer was run, this data could be used with other information to attribute the result to this change.
In all cases, operationalizing forecasts is a hybrid effort between the team and the machine, blending the automation of otherwise manual work with the knowledge and expertise of those onsite who know visitors best.
Get insights delivered right to your inbox
Want to learn more about Dexibit?
Talk to one of our expert team about your vision to discover your data strategy and see Dexibit in action.
