Prädiktive Besucherfrequenz und KI: Prognose von Besuchermustern aus WiFi-Daten

PODCAST SCRIPT: Predictive Footfall and AI — Forecasting Visitor Patterns from WiFi Data Duration: ~10 minutes | Voice: UK English, Senior Consultant Tone --- [SEGMENT 1 — INTRODUCTION & CONTEXT — approx. 1 minute] Welcome. If you're responsible for a venue, a retail estate, or a hospitality operation, you've probably been told that your WiFi network is sitting on a goldmine of data. And that's true — but only if you know what to do with it. Today we're going to talk about predictive footfall analytics: what it actually means in practice, how the machine learning works, what data you need to make it reliable, and — critically — how organisations are using these forecasts to drive real operational decisions right now. This isn't a theoretical exercise. The organisations getting the most value from WiFi-derived footfall forecasting are using it to cut staffing costs, reduce stock waste, and time their marketing pushes to within the hour. That's what we're here to unpack. --- [SEGMENT 2 — TECHNICAL DEEP-DIVE — approx. 5 minutes] Let's start with the data layer, because this is where most implementations either succeed or fail before they've even begun. Your WiFi infrastructure — whether that's a managed network running 802.11ax access points or an older 802.11ac estate — is continuously collecting probe requests and association events from every device in range. Each of those events carries a timestamp, a signal strength reading — that's RSSI, Received Signal Strength Indicator — and, historically, a device MAC address. Now, MAC address randomisation, introduced aggressively from iOS 14 and Android 10 onwards, has complicated device-level tracking. But here's the thing: for footfall forecasting, you don't actually need persistent device identity. You need aggregate counts, dwell time distributions, and zone transition patterns. Anonymised, aggregated data is both GDPR-compliant and entirely sufficient for the forecasting models we're going to discuss. So what does the data pipeline look like? At ingestion, your access points are streaming probe and association events to a central controller or cloud platform. The pre-processing layer handles deduplication — because a single device will generate dozens of probe requests per minute — and applies anonymisation. From there, feature engineering extracts the metrics that actually feed the model: hourly visitor counts per zone, average dwell time, entry and exit rates, and crucially, external covariates like day of week, public holidays, local events, and weather data. Now, the model selection question. This is where I see the most confusion in the market. Organisations either default to simple moving averages — which are essentially useless for anything beyond a 24-hour horizon — or they jump straight to deep learning without the data volume to support it. Here's a practical framework. If you have six months of clean hourly data and your venue has relatively stable seasonal patterns — think a commuter-facing coffee shop or a supermarket — SARIMA, that's Seasonal AutoRegressive Integrated Moving Average, will give you solid 7-day forecasts with mean absolute percentage errors in the eight to twelve percent range. That's good enough to drive staffing decisions. If you have twelve months or more and you're dealing with irregular spikes — concerts, bank holidays, promotional events — Facebook's Prophet model is worth deploying. Prophet handles changepoints and holiday effects natively, and it's interpretable enough that your ops team can understand why the model is predicting a surge on a given Saturday. For venues with rich feature sets — a large retail estate where you're feeding in promotional calendars, competitor activity, and loyalty programme data alongside the WiFi signals — gradient boosting models like XGBoost consistently outperform statistical approaches. With twelve months of training data and good feature engineering, you're looking at mean absolute percentage errors in the three to six percent range. That's the level of accuracy where you can genuinely automate stock replenishment triggers. And then there's LSTM — Long Short-Term Memory neural networks. These are powerful for capturing long-range temporal dependencies, but they need eighteen months of data minimum to train reliably, and they're computationally expensive to retrain. I'd recommend LSTM for large-scale deployments — think multi-site retail chains or stadium operators — where you have the data volume and the engineering resource to maintain the model. One thing that catches organisations out: the difference between a WiFi-connected visitor count and a true footfall count. Not every visitor connects to your WiFi. Capture rates vary enormously — from around thirty percent in a quick-service restaurant to over eighty percent in a hotel lobby where guests are actively seeking connectivity. You need to calibrate your WiFi-derived counts against a ground-truth source — door counters, POS transaction volumes, or manual counts — before you can trust the absolute numbers. The relative patterns — the peaks, the troughs, the day-of-week rhythms — are reliable almost immediately. The absolute counts need that calibration layer. On the infrastructure side, access point density matters more than most people realise. For zone-level footfall granularity — meaning you can distinguish between different areas of a floor — you need access points no more than fifteen metres apart, with overlapping coverage cells. This isn't just about connectivity performance; it's about triangulation accuracy for the positioning layer that feeds your zone-transition data. The Indoor Positioning System guide on the Purple blog goes into the technical detail on UWB, BLE, and WiFi-based positioning if you want to go deeper on that. --- [SEGMENT 3 — IMPLEMENTATION RECOMMENDATIONS & PITFALLS — approx. 2 minutes] Let me give you the three things that determine whether a predictive footfall deployment actually delivers ROI, or ends up as an expensive dashboard that nobody looks at. First: data quality over model sophistication. I have seen organisations spend six months selecting and tuning an LSTM model on dirty data, when a well-calibrated Prophet model on clean data would have delivered better forecasts in six weeks. Invest in your data pipeline first. Specifically: get your deduplication logic right, handle MAC randomisation with session-based counting rather than device-level tracking, and establish your calibration baseline against a physical count source before you touch a model. Second: define the downstream decision before you build the model. The forecast is worthless unless it's connected to an action. The most successful deployments I've seen start with the operational question — "how many staff do I need on the floor at 2pm on a Tuesday in December?" — and work backwards to the model specification. That determines your forecast horizon, your granularity, and your acceptable error tolerance. A staffing decision needs a 7-day forecast at hourly granularity. A stock replenishment decision for a distribution centre might need a 14-day forecast at daily granularity. Those are different models with different data requirements. Third: plan for model drift. Visitor behaviour changes. A new competitor opens nearby, a transport link closes, your venue undergoes a refurbishment. Models trained on pre-change data will degrade. Build a retraining cadence into your operational process — monthly for most venues, weekly if you're in a high-volatility environment like events or transport hubs. The GDPR angle is worth flagging explicitly. WiFi-derived footfall data, when properly anonymised and aggregated, does not constitute personal data under the UK GDPR or EU GDPR. You are not tracking individuals; you are counting devices. But your privacy notice should still reference the use of WiFi signals for venue analytics, and you should ensure your data retention policies cover the historical training data you're holding. --- [SEGMENT 4 — RAPID-FIRE Q&A — approx. 1 minute] Let me run through the questions I get asked most often. "How much history do I actually need?" Minimum six months for a useful SARIMA model. Twelve months to capture a full seasonal cycle. Eighteen months if you're going LSTM. "What accuracy should I expect?" For a well-implemented XGBoost model with good features, three to six percent MAPE on a 7-day horizon is achievable. For simpler models on shorter horizons, eight to twelve percent is realistic. "Can I use WiFi data alone?" Yes, for relative pattern forecasting. For absolute count forecasting, you need a calibration source. "What's the minimum AP density for zone-level analytics?" One access point per 150 to 200 square metres for basic zone counting. One per 80 to 100 square metres for reliable dwell time and transition data. "How long does a full deployment take?" Eight to twelve weeks from data audit to first production forecast, assuming clean infrastructure and a defined use case. --- [SEGMENT 5 — SUMMARY & NEXT STEPS — approx. 1 minute] To summarise: predictive footfall analytics from WiFi data is mature technology. The models work, the accuracy is sufficient for operational decisions, and the ROI is demonstrable — typically in staffing efficiency and stock optimisation within the first quarter of deployment. Your immediate next steps: audit your existing WiFi infrastructure for data completeness — are you logging probe and association events? Establish your calibration baseline. Define the operational decision you want to automate or improve. And select your model based on your data volume, not on what sounds most impressive. If you're running Purple's WiFi Analytics platform, the data pipeline and anonymisation layer are already in place. The question is whether you're using the historical data you're already sitting on to drive forward-looking decisions, or whether you're still looking at last week's dashboard. That's the difference between reactive analytics and predictive intelligence. And that's where the real operational value lives. Thanks for listening. Links to the full technical guide, architecture diagrams, and implementation checklist are in the show notes. --- END OF SCRIPT Total estimated duration: ~10 minutes at 140 words per minute (script is approximately 1,380 words)