Five Mistakes I Made Building My First Predictive Analytics Dashboard in Perth

Last summer, a dashboard graph on my office monitor was pointing in a completely wrong direction. The customer churn rate was showing as negative. The developer sitting next to me glanced over and couldn't hold back a laugh. I took a sip of coffee and started tracing back where things went wrong.

Taking charge of predictive analytics while handling marketing at a Perth startup was purely stubbornness. I turned down the suggestion to hire an outside consultant and started from the design stage myself. That stubbornness did lead to learning, but I took plenty of detours along the way.

I designed the dashboard before collecting the data

The very first thing I did was lay out the visualisations. I opened Tableau and spent two days figuring out chart arrangements. Meanwhile, I hadn't even sorted out what data to pull or where to pull it from.

• Customer data from the CRM and web log data were in different formats
• One date field was YYYYMMDD, the other was DD/MM/YYYY
• Hundreds of rows went missing during the merge, and I didn't notice until much later

I should have designed and validated the data pipeline first. Pretty charts only mean something when they sit on top of clean data.

I crammed in twelve KPIs

I wanted the dashboard to cover everything. Churn rate, conversion rate, CAC, LTV, session duration, email open rate — the lot. With the screen packed full of numbers, team members stopped looking at the dashboard in meetings altogether. Pure information overload.

I narrowed it down to the three metrics the CEO wanted to see and the two I actually used to adjust campaigns — five total. Only then did the dashboard become the centre of conversation in meetings.

A model that ignored Perth's market characteristics

I used benchmark data based on Sydney and Melbourne as-is. Perth has a different population size and different consumer behaviour patterns. Applying east coast seasonal trends directly made the prediction accuracy dismal. Perth's summer spending patterns in particular were markedly different from the east coast, and failing to account for that in the model was a clear mistake.

I should have collected at least six months of local data separately and set up region-specific variables.

Photo: Orderinchaos, CC BY-SA 4.0, via Wikimedia Commons

I tried to do everything myself

I'm the type who makes a plan and pushes through no matter what. This time, that worked against me. Writing Python scripts, cleaning data, building visualisations, training the team — I took it all on. About two months in, burnout hit, and as updates fell behind, there were stretches where the dashboard was stuck showing data from two weeks prior.

I should have automated what could be automated and handed the data engineering off to the dev team. A marketing manager's job is to extract insights and connect them to strategy, not to maintain pipelines.

When the predictions were wrong, I didn't question the model

When the dashboard showed "30% increase in leads next quarter," I put that number straight into the executive report. The actual result was only a 10% increase. The model's underlying assumptions — that the ad budget would hold steady, that competitors would maintain the status quo — had broken down in reality, and I never checked.

A predictive model is a directional indicator, not a definitive number. After that, I made it a habit to present ranges by scenario.

---

That dashboard is now on its third version. It's still not perfect, but the team actually opens it every week and uses it for decision-making. That should have been the goal from the start — not a pretty graph, but a useful tool.