The Journey is your builder’s reference library. Crack open the full archive here → https://6catalysts.substack.com/archive
Folks who call economics the ‘dismal science’ probably haven’t struggled through building a quarterly sales forecast in a spreadsheet. For many builders, sales forecasting (“demand forecasting”) has something in common with public speaking.
Sweaty palms.
Like public speaking, most people dread it until they’ve done it enough times to realize they won’t actually die on stage. The same is true for sales forecasting.
If you’re making tangible things, anyway. More often than not, there’s less at stake with forecasting demand for an early-stage company that’s operating in services or intangible goods. So, we’re going to focus on this topic today through that lens—creating and producing tangible goods.
What is Forecasting?
Simply put, it’s making a bet on how many things your future customers will buy. You use that forecast to go out and place orders with suppliers. If you use a contract manufacturer (or co-packer) to produce your things, you’ll usually give them a forecast for finished goods.
Forecasting is an often under-appreciated function within a scaling company. Having a forecasting methodology in place that consistently and accurately estimates demand is really fucking hard. Even giant companies with decades of experience and significant resources routinely get it wrong. This is one of the reasons why overstock retailers like Winners/Marshalls (TJX Group) and Liquidation World (LW Stores) exist with viable business models.
There are three basic categories of forecasting, with a number of distinct methods in each :
Causal modelling focuses on the impact of cause-and-effect factors on demand (for example, macro trends like “how’s the economy doing” and micro trends like “we’re running a sale this month”).
Qualitative modelling focuses on building a supporting picture of the demand landscape beyond statistical prediction (for example, market research or research on brand affinity / loyalty).
Time-series modelling focuses on using your own demand data to predict future demand
Causal modelling and qualitative modelling are often complex endeavours that are of limited use to new ventures and small-but-scaling organizations (with the exception of price-elasticity calculation, which is a type of causal modelling).
So for today, we’re going to focus on Time-series modelling. It’s accessible to practically any organization, and if you have some basic spreadsheet skills, you can benefit from it almost immediately. Once you begin to grow and can invest in systems and tools, there are a number of great forecasting tools available that you can use to strengthen your muscles in this area.
Why does Forecasting Matter?
Optimization of resource utilization : specifically, your time and money.
So, forecasting is intimately tied to the health of the business… and more than most functions of running a business, this is a Goldilocks thing. Not too hot, not too cold, but just right. Easier said than done.
Some leaders will argue if you’re going to miss the mark, it’s better to slightly under-forecast so that there’s more stability in your cash conversion cycle. But, you leave money on the table since you’re out of stock.
Some leaders will argue that it’s better to have slightly over-forecasted so that stock outages don’t constrain sales results. But, you tie up cash which could be used for other things (like investments in the business).
Personally, I think it’s a moot point. Obsessing about the over/under on forecasting is counterproductive, and can distract you from executing. Base your forecast on a rational estimate of demand using the data that you have, apply some guardrails to prevent financial calamity in the worst case scenario, and get to work.
How to Forecast Demand
Every company will do sales forecasting a little bit differently. Because you’re essentially creating an (educated) guess on the future, there’s no one-size-fits-all way to do it.
With that understanding, you should feel empowered to experiment with different methodologies and processes—and fine-tune them over time. One thing that many companies don’t do is track the effectiveness of their forecasting over time and evaluate what works and what doesn’t. Don’t be one of those companies. Continuous improvement pays you back compounding dividends on your investment well into the future.
Predictive Modelling Methods for Sales Forecasting
For each predictive model, I’ll include an example in a linked Google Sheet. Feel free to copy the sheet to your own Google Account (or download it) to play around, or use during your own forecasting adventures.
Time Series : Historical Comparison
This is the simplest method, but it also lacks nuance and doesn’t account for deviations and growth through marketing or other activities that increase your unit sales volume.
In a historical comparison, you really just look back to a comparable time period and assume that you’ll sell the same amount of things in a given future period.
Sold 228 widgets in January of this year? Great, you’ll also sell 228 widgets in January of next year.
You can see the problem with this model, I’m sure.
Though on its own, historical comparison is very limited (and barely better than guessing), it forms the basis for all other models. It’s one of the few data points that’s definitively known when you create your forecast.
Time Series : Moving Average (MA)
The moving-average modelling method slightly de-risks your forecast by smoothing out volatility of demand across time periods. Adding in this smoothing factor can improve the accuracy of a forecast by flattening the impact of noise (unusual demand or weakness within a given time period).
To upgrade your historical comparison with a moving average, do this :
Gather a full year’s unit sales data, broken out by relevant time periods (monthly is quite common, but other time series are more appropriate in some cases)
Calculate the average of three times’ the number of time periods for each time period (i.e., calculate the average of three months if your time series is monthly)
Apply your 3x period average as the forecast for each individual time period
Time Series : Seasonally-Adjusted Moving Average (SAMA)
Adding seasonality adjustments is another layer that can improve the accuracy of your forecasting—many of these models can be combined in a complementary way. A seasonal adjustment captures trends that are predictable for a given period of time. For example, if your business sells consumer goods and you traditionally see an uptick in demand around public holidays, you can adjust your moving average to factor in this activity.
To upgrade your historical comparison with a seasonality adjustment, do this :
Identify any recurring seasonal periods where your business has, or likely will, see extra-ordinary peaks or valleys in customer demand (for example, during gifting holidays like Christmas)
Gather a full year’s unit sales data, broken out by relevant time periods (monthly is quite common, but other time series are more appropriate in some cases)
Calculate the average of three times’ the number of time periods for each time period (i.e., calculate the average of three months if your time series is monthly)
Apply your 3x period average as the forecast for each individual time period
Using your historical data, calculate the percentage lift or drop in unit sales between the seasonal time period and the previous time period. This is your seasonality factor.
Adjust the moving average for each seasonal period by its seasonality factor.
If you’ve elected to use a year-over-year historical comparison, then there’s limited value in applying a seasonality adjustment on top of your moving average—in this scenario, the seasonality factor is already an inherent part of the model (though it will be a bit more muted than if you actively apply an adjustment).
If you’ve elected to use a quarter-over-quarter or month-over-month historical comparison as your base model, then a seasonal adjustment can be very useful at capturing that extra-ordinary demand (or lack thereof).
Time Series : Exponentially-Smoothed Bursting
This model can help improve forecasting accuracy for organizations which rely on intense pulses or bursts of promotional activity to generate demand (quite common for very new businesses). It can also be very useful for forecasting demand around a new product launch.
Many established businesses struggle to forecast demand for new products, as there isn’t a set of historical data to draw from (we’ll talk about this a bit more later in this post). I developed this model specifically for that use case. It’s a bit of a bastardization of the moving average and a statistical method called exponential smoothing.
To upgrade your historical comparison with this model, do this :
Identify historical periods where your business produced a specific “burst” of sales tied to an active, focused effort (for example : promotional/PR activities, product launches)
Gather a full year’s unit sales data around those burst periods, and break them out by relevant time periods (monthly is quite common, but other time series are more appropriate in some cases)
Calculate the average period unit sales for a year following each burst (if possible). This is your baseline.
Calculate the delta in unit sales from the baseline for each period leading up to and following the burst period(s).
Use the deltas for each period to create a demand curve, that can be applied to model the impact of a future burst.
Similar to a seasonality-adjustment, apply the deltas from your demand curve as a modifier to the unit sales for the periods after your burst period. For example, period one might be +90% (*1.9), period two might be +60% (*1.6), et cetera.
This method creates a measured window that will help you model an unusual demand curve that wouldn’t otherwise be well-captured by using the models above. This model works best when applied to forecast a specific product or group of products that are highly similar (it’s considerably less effective when applied indiscriminately to product groups which aren’t similar).
What if You Don’t Have Historical Data?
This category of predictive modelling really relies on the certainty of past results to generate confidence in future results (the ‘educated guess’). Weak, iffy, or a complete lack of historical data to draw from really hollows out the effectiveness of this category of predictive modelling. So, what can we do if we don’t have highly-relevant historical data to draw from?
In the case of a new product launch, you won’t have historical data for that product (obviously). If you have similar products, you can draw from the historical data of the closest product or product grouping. You might look at factors like typical use cases, target customer, and target price point in deciding which similar product to use.
In the complete absence of historical data (which will happen if your business is brand new), it can be useful to instead create your forecast based on financial risk.
For example, let’s say that you have $10,000 in operating capital that could go toward production (add zeroes to that number if you need to). Ask yourself a few questions :
How much would I need to hold in reserve for another production run if this one was a complete disaster and I sold nothing?
How confident am I that I’ve found product/market fit?
If I needed to reduce my profit margin through promotions to improve unit sales, how much would that compromise my ability to try again?
Essentially, you base your forecast on your tolerance for risk given the things that you know. If you have $10,000 in operating capital and trying again would cost you $5,000, then you might produce as many units with the other $5,000 as you could. Go to market, measure your results, adapt, and go again.
In isolation, basing a forecast solely on your own tolerance for risk can be a bit unnerving. Unfortunately, the emotional ups and downs of that are a part of the game. But, you can build confidence in other ways through rapid prototyping and customer insights research before you commit to a full production run.
Recently, I wrote about a few methods that were developed to validate ideas—but they can also be applied to product development and prototyping (link below).
The Spark Is Not the Strategy
One of the most-romanticized characteristics of entrepreneurs and builders is this :
Feedback
Improving The Journey is a journey in itself—and I’d love your help. Share your anonymous feedback here : https://6catalysts.ca/subscribe/the-journey-feedback/
