The difference between good ecommerce brands and exceptional ones often comes down to how confidently they make decisions about the future.
You can get lost in the woods of marketing metrics but lose sight of the bigger picture, and the ultimate goal: profitable revenue growth.
The best in the game know how much to spend and when, which channels will deliver the best ROI, and how aggressive they can be with discounting without eroding long-term customer value. They also acknowledge what they do not know and have a solid roadmap to get smarter.
The answer lies in causal forecasting.
What is Causal Forecasting for Marketing?
At its core, forecasting uses historical data to predict what will happen in the future based on the business or marketing decisions you make today.
Combining models of new customer acquisition and customer retention helps brands make the best decisions, depending on their business objectives.
Unlike basic trend analysis or spreadsheet projections, causal forecasting is rooted in causal inference.
What is ‘causal inference’: Causal inference is a branch of data science that identifies whether one thing causes another, not just whether they move together. It separates true cause-and-effect from misleading correlations by accounting for confounding factors, timing, bias, and real-world mechanisms. In marketing, causal inference helps answer questions like: “Did this campaign actually drive incremental sales?” rather than “Did sales go up while I was running ads?”
The goal is to model meaningful strategic decisions, not simply predict the future well if things stay the same. This is only possible if the data science is driven by real marketers who combine domain knowledge with statistics to choose optimal strategies, estimate risk and guide decisions. The alternative (pushing data through a “one size fits all” algorithm or tool) fails to capture the nuance of real-world business decisions. This approach isn’t about chasing model fit statistics; it’s about using rigorous data science methods to support decisions that actually move the needle.
Forecasting can answer questions like:
- How many customers will we acquire in Q4 based on our planned spend and discount strategy?
- What happens to revenue if we increase Meta spend by 30%?
- How much revenue will we generate within 3 months from customers acquired at discount versus full price?
- If we dial spend up or down, how will this affect new and returning customer revenue?
- What EBITDA% target will drive the most long-term growth?
- What should CAC be throughout the year?
Why Forecasting Matters More Than Ever
In the wake of privacy initiatives like iOS14 and the deprecation of third-party cookies, more marketers than ever have been waking up to the reality that they’re flying blind. A surge of services and tools are flooding the market, claiming they can measure everything through modelling. Many of these solutions capitalise on the marketing industry’s lack of deep scientific expertise, promising certainty where none exists.
Pixel measurement has always been fundamentally broken. Attribution models consistently undervalue awareness-driving activity, making it look like revenue is coming from nowhere, “organically”.
This creates a dangerous trap where brands achieve seemingly impressive blended efficiency metrics whilst massively underspending on upper-funnel channels during high-demand periods. They’re missing opportunities to acquire valuable customers who would generate repeat revenue for years to come, because their measurement systems don’t properly credit the awareness-building activity that drives discovery and consideration.
Causal forecasting cuts through this noise. By modelling the causal relationships between marketing activity and business outcomes, it can reveal the real levers of growth, even when those levers don’t show up clearly in your GA dashboard. The goal is to predict the end result accurately, so when something works, it will be obvious.
The Real Business Problems It Solves
Predictive forecasting is not an academic exercise. It is a practical tool that solves the thorniest challenges facing ambitious ecommerce brands:
Set better Targets
What is the best CAC target, given the overall business growth goals? Should these change over time? What levels of spending on marketing will give us the best balance of business growth and efficiency?
Plan for peak season
Know exactly how much to spend during peak season to maximise customer acquisition without destroying margins. A leading alcohol brand we worked with grew revenue over 40% YoY in Q4 by switching from micro-control with platform metrics to our forecasting and measurement system.
Allocate budget across channels
Stop guessing which channels deserve more investment. We helped an online furniture marketplace diversify away from heavy reliance on Google Search to include paid social. Multi-touch attribution (MTA) and platform results would have told them to avoid the channel entirely, but forecasting revealed the true value of upper-funnel activity.
Create a profitable promotion strategy
Understand the long-term impact of discounting. That 25% off promotion might look great in the short term, but what about 90-day, or even 365-day customer LTV? Forecasting helped one of our clients discover their promotion was not as effective as they thought because the acquired customers had lower long-term value.
Make better Stock and Inventory Decisions
Plan stock levels with confidence, knowing what demand will look like months in advance, based on your marketing plans.
Board Reporting and Investment Planning
Create credible business plans for investors with scenario modelling that quantifies risk and opportunity.
How It Works: The BARK Approach
Our forecasting methods have evolved directly from our clients’ needs as ambitious ecommerce businesses. We do not put datasets through a fixed process and present whatever comes out. Our approach combines:
Time Series Regression Modelling
Understanding how performance changes over time, accounting for seasonality and trends.
Media Mix Modelling
We use our own custom frequentist MMM process (a statistical approach that estimates channel contributions using fixed parameters rather than probability distributions) alongside Meta’s Robyn and Google’s Meridian. Our models can incorporate interaction effects and bake these into forecasts, which is essential for understanding how channels work together. This capability earned us recognition as the only UK paid media and creative agency to also be accredited for measurement by Meta.
Causal Inference
Why Causal Inference Matters
The critical difference. Everyone has access to advanced algorithms, but understanding what will happen when you make changes requires expertise in causal inference. There is no tool that can magically do this. It requires expert analysts who understand marketing, and communicate how certain we can or cannot be about what is driving revenue. Whilst your average data scientist is laser focussed on making models with the best fit to a dataset, our methods are derived from the field of epidemiology, where the goal is to estimate the risk for diseases, and understand their causes.
The Challenge of Synced Spend Patterns
Here’s a common challenge in MMM: when marketing spend across different channels rises and falls in sync (as often happens with seasonal budget allocation) there is little evidence for what will happen if you change that pattern. Some MMM tools will produce models that forecast beautifully, but these same tools may have also considered alternative models that forecast equally well yet assign dramatically different importance to each channel. Without proper causal inference techniques, you might be looking at a model that fits the data perfectly but would fail spectacularly if you actually changed your strategy based on its recommendations.
Why Experiments Alone Aren’t Enough
Using experiments to calibrate MMM models can help make them more valid, but this alone doesn’t make a model “causal”. There is still so much more to it. True causal inference requires transparency about uncertainty. Rather than claiming every model is perfectly causal because it follows a standard process, the best approach acknowledges that estimating causality is inherently difficult. It requires close collaboration with clients and a custom, human-centred methodology. No model perfectly captures causality, but recognising this limitation and working to address it, is what separates rigorous forecasting from oversimplified prediction.
We are honest about uncertainty. We explain our models’ limitations thoroughly. Making data-led decisions is not about blindly following maths, it is about using data to help make decisions and take calculated risks.
What You Need to Get Started
To build robust forecasts for new customer acquisition, you typically need:
- Daily spend on marketing channels
- Dates or impressions for other marketing, affiliate, and influencer campaigns
- Daily new customer sales
- All store order data, including promotion codes and discounting information
- Dates and information on PR activity, promotions, stock-outs, and other contextual factors
The good news is there are no hard prerequisites. Forecasting is almost always possible to some extent. The depth of history and volume of data simply influence how robust the results will be and what scenarios we can usefully simulate. The richness of contextual information is the most important factor here.
Best Practice: From Data to Decision
Forecasting shouldn’t be a one-and-done deliverable: it’s a continuous journey towards better decision-making.
The best forecasting engagements begin the moment data is available, with immediate feedback on dataset quality and potential gaps. From there, models are built, hypotheses are tested, and insights are continuously refined based on real-world results.
Some models retrain daily to capture the latest patterns. Others remain fixed after major forecasting projects to track how predictions hold up over time. Either way, daily feedback loops show whether results are unfolding as expected and whether long-term forecasts need adjustment.
Deliverables should match business needs: dashboards that track performance against predictions, Looker Studio or presentations linked to Google Sheets for hands-on scenario testing, or custom web apps where you can adjust variables like CAC targets or future budgets and instantly see predicted outcomes.
Scenario Modelling: Planning for Multiple Futures
One of the most powerful applications of forecasting is scenario modelling. Rather than making a single prediction, this models multiple futures:
Scenario A: Take a hit on margin during peak demand to acquire a large volume of customers.
Scenario B: Keep margin healthy throughout the year, spend less, and acquire fewer customers.
Depending on expected lifetime value and how customer revenue compounds over time, one scenario will be clearly better for maximising growth at your desired margin. The forecast shows you which.
This kind of analysis is invaluable for annual planning, board presentations, and any situation where you need to quantify the long-term impact of short-term decisions.
Validating the Forecasts
We are asked all the time about forecast accuracy. The truth is it is not a competition to achieve the highest accuracy percentage. It is an endeavour to understand the levers of business performance and quantify risk.
That said, any forecast should be validated rigorously, using:
- Robustness Checks: Changing the training date range slightly should not massively alter model parameters.
- Hold-Outs: Models should be trained blind to recent results and tested on predictions. Critically, the focus should be on counterfactuals: testing that models predict correctly when something changes, not just when everything stays the same.
- Real-World Experiments: The ultimate validation. Models suggest something, teams test it, and the hypothesis is proven or disproven.
Strong forecasting requires close collaboration between analytics teams and business stakeholders. If everyone in your business knows last week’s sales were great because the brand was mentioned on TV, but the forecasting team isn’t aware, that’s a massive issue. Business context must be baked into modelling every step of the way.
When Forecasts Get It Wrong (And Why That Matters)
It is important to be honest: forecasts have limitations.
A model might have 95% accuracy on hold-out data, but the universe is not obliged to continue drawing the same curve of diminishing returns when you double your spend. Many businesses are highly seasonal. If that seasonality is not modelled correctly, you will get sub-optimal budget recommendations.
We have seen brands test new channels whilst running more enticing offers. The MMM thought the new channel was brilliant, because it did not know about the offer. We have seen brands use MTA-based forecasts built on low-season data, only to watch assumptions fall apart when high-season demand changes conversion behaviour dramatically.
This is exactly why expertise in causal inference matters. It is not enough to build models that fit the data. You need to understand when those models can be trusted to guide decisions and when they cannot.
If the evidence isn’t strong enough to determine whether campaign A is better than B, your agency should tell you exactly that (not fudge the statistics). They should help you plan the right experiments to learn what you need with confidence.
Who Benefits Most?
Honestly? Everyone.
But predictive forecasting becomes especially valuable for:
- Brands with complex seasonal patterns that make it hard to separate marketing impact from natural demand fluctuations
- High-growth brands testing new channels and needing to understand true incremental value
- Brands with repeat purchase models where understanding long-term customer value is critical to profitability
- Established brands with mature channel mixes who need to optimise budget allocation across multiple platforms
If you are experiencing stagnated growth, failed tests, or heavy uncertainty about what is working across your marketing strategy, predictive forecasting moves from “nice to have” to “essential”.
The Cost of Not Forecasting
The most common mistake we see is brands underspending on upper-funnel marketing during high-demand periods because their measurement systems undervalue awareness-driving activity.
It looks like you are hitting great efficiency targets. Your board is happy. But you are actually missing massive opportunities to acquire customers who would deliver years of repeat revenue. By the time you realise it, the high-season window has closed.
Other brands overspend on channels that look good in isolation but do not actually drive incremental growth. Or they make promotional decisions that seem smart in the moment but erode long-term customer value.
These are not small mistakes. They are strategic errors that compound over time, costing millions in missed growth or wasted investment.
Getting Started
Predictive forecasting does not require a massive commitment. We work with brands on everything from one-off projects to ongoing partnerships as part of our integrated growth package.
For some brands, we start with a scoping project, reviewing existing MMM work or assessing how much insight we can draw from current datasets. For others, we dive straight into building comprehensive forecasting models as part of their broader growth strategy.
Either way, the journey starts the same: understanding your key business goals, exploring your data, and learning the history of your business and marketing strategies.
The question is not whether you can afford to invest in predictive forecasting. It is whether you can afford not to, whilst your competitors are already using it to plan their next move.
