The Modeling System That Powers 30%+ Revenue Growth

[00:00:00] Luke: Hey everyone, welcome to our Canon series. This is part two of a series that we've been going through that's walking through CTC's methodology Canon across the core dimensions of how we approach forecasting, modeling, measurement, creative strategy, meta advertising, Google advertising email strategy, and on down the line.

[00:00:23] So this is part two, where we're focusing on modeling and how we approach modeling that sits at the center of our operating system. This is coming after part one, which is around forecasting and our approach to that, which is a good place to start. But I'm going to dive into the Canon methodology around modeling and why this is so crucial and central to how we do things at CTC.

[00:00:46] So there are four proprietary models that power CTC forecasting and planning system. Each model answers a s- a specific question about the future of the business, and together they produce the daily P&L forecast that drives every operational decision. So four models powering what we covered in the forecasting section, which is three point one percent forecast accuracy across over three billion in, in GMV managed for us producing over thirty percent revenue growth, over forty percent contribution margin growth across our data set as well, with the goal of being daily posi-precision, every dollar, every day accounted for and with a target that allows us to execute against the forecast and make it a reality.

[00:01:28] So before we dive into each model, it's important to understand how they work together. These four models form a connected system where each output feeds the next. The spending power model answers the first question. Given a high level of spend, how efficiently will that spend convert into new customer revenue?

[00:01:46] It produces a spend to effi- spend to efficiency ge- degradation curve that tells the, the profit engineer, the operator sitting at the intersection of the system exactly how many new customer dollars to expect at any budget level. Those new customers then flow into the retention model. For every cohort of new customers acquired in a given month, the retention models predict how much revenue they will bring back in each subsequent month.

[00:02:06] This output forms a complete picture of what the returning customer revenue built from the bottom up, bottom up is cohort by cohort. The third model is the event effect model, which adds daily precision, precision. So on any given day, a marketing event, a sale, a product drop, a VIP campaign shifts the natural pattern of revenue.

[00:02:25] The event effect model learns these shifts from a historical data and then applies them to future event days on the marketing calendar. It does this without breaking monthly budget tar- targets. Event days run harder, surrounding days compress proportionally. The month always reconciles to what the monthly target is set again by the spending power model and the retention model.

[00:02:44] Then finally, the creative demand model ensures the ad account has enough fuel to execute the plan. Based on the brand's current ad portfolio health and the planned spend level ca- it calculates exactly how many new ads need to be produced each month. This closes the loop. The business plan determines the spend, the spend determines the creative demand, and the creative, creative enables that spend.

[00:03:02] So the four models, spending power model to forecast new customer revenue given any level of spend. Re-retention model forecasts returning customer revenue from every past cohort. Event effect model distributes revenue accurately across days with marketing events. And the creative demand model ensures the ad account has enough creative fuel to sustain the plan.

[00:03:20] Together, these form the model system that are connected to produce a daily P&L level forecast, where we have every dollar every day mapped out and can execute against it to make that forecast a reality. Because as we know, forecasting is as much an exercise in execution as it is in planning. So we're gonna deep dive into each one of these models to give some more context into what goes into building the models and how critical they are to a system that's producing predictable profit.

[00:03:53] So the first model to deep dive into, the spending power model The model is built on ACoNS, which is ad cost and new sales, defined as total ad spend divided by new customer revenue. This is a new customer-focused lens on ACoS or ad to sales, ad cost to sales and is the inverse of AMER, acquisition marketing efficiency rating.

[00:04:14] It represents how many cents of ad spend it takes to generate one dollar of new customer revenue. Lower ACoNS means higher efficiency. The model produces a linear equation for each month, a slope and an intercept that describes how ACoNS changes as spend increases. The slope captures diminishing returns.

[00:04:32] The more you spend, the less efficient each incremental dollar becomes. The intercept represent- represents baseline efficiency at near zero spend. A related output in the model is spending power, which is defined as the inverse of the slope, one divided by the slope. So spending power tells you how many additional dollars of monthly spend it takes to degrade ACoNS by one point.

[00:04:57] A million dollars is high spending power, right? The brand can absorb a million dollars of incremental spend before efficiency drops meaningfully. Strong creative, broad TAM, efficient machine, et cetera. Two hundred thousand dollars may indicate low spending power. Efficiency falls off quickly with scale.

[00:05:11] Narrow audience, creative fatigue, or market saturation. Scale requires caution and creative investment. These all form the ACoNS curve which for every brand is a unique degradation curve that has a differing level of how steep that curve is related to the spending power that it represents, right?

[00:05:30] So the brands that are strongest, that have strong spending power, have a flatter curve than those with low spending power, where it falls off a cliff and degrades much more quickly, meaning that the incremental ten thousand dollars in spend, twenty thousand dollars in spend, et cetera, degrades the acquisition efficiency more meaningfully for that brand.

[00:05:51] So how the profit engineer uses the spending power model is that in the Statlas planning module, the profit engineer inputs a spend plan for a future month, and the model applies that month's forecast slope and intercept to compute expected ACoNS from which AMER, new customer revenue, new customer account, and CAC are all derived.

[00:06:11] This is the first building block of the monthly revenue forecast. This model also helps us to establish valid spend ranges based on historical data. So predicting efficiency at spend levels well outside the historical range should be treated with caution. Said another way, we use this as the best reflection, this model is the best reflection of what is likely to occur, and when we start to manually adjust it in either direction we need to understand that there are bounds in which the model exists, and it represents new activity that is going to be necessary to achieve that outcome outside of any historical baseline that it represents.

[00:06:47] The spend efficiency curve isn't just a forecast, it's also a dec-decision tool. So there are three points on the curve that we plot clearly in terms of the business optimization against the spend allocation. So, the three optimizations that we focus on around the spend AMIR model is to maximize month one contribution margin is the first scenario.

[00:07:07] So I want the maximum amount of n-new customer contribution margin dollars from my new customers. The second scenario is new customer revenue at break even. So I want to maximize the new customer revenue I'm generating at a break even contribution margin outcome. And the third optimization being maximize li-lifetime contribution margin.

[00:07:26] So this is similar to the first scenario in which my goal is to maximize contribution margin for my new customers, but I'm willing to wait a longer-- against a longer time horizon to realize that value. So I'm willing to wait sixty or ninety or a hundred twenty days and on down the line, a set defined LTV time period to allow for increased volume within month one and to capture the latent impact of those customers along a broader time horizon and maximize the contri-contribution.

[00:07:54] So each of these points, max CM, max revenue, max lifetime value CM, represents a really fundamentally different business strategy. A cash-constrained brand may need to optimize for a month one contribution. A brand with strong retention may push towards lifetime contribution margin and even accept short-term losses for long-term customer file growth.

[00:08:14] The model makes these trade-offs explicit and computable. The P does not-- The profit engineer on our side isn't guessing which point to target. The brand isn't guessing which point to target. The model contains each scenario in Statlas and shows the business the financial implications of each choice and what the trade-offs are to align it with the core business objective.

[00:08:33] So every model for the spending power model shows a spending power trend over time, predicted versus actual accuracy metrics across all historical forecast demands, platform by platform allocation recommendations, incrementality estimates per platform, and a care-- in comparison a brand's trend against CTC's cross-brand direct consumer, consumer index as well to understand where the brand sits relative to the broader data set That is model one, the spending power model.

[00:09:01] The second model is the retention model, which is all focused on forecasting returning customer revenue. It answers, given the new customers we acquired each month, how much revenue will they bring, bring back in future months? The core metric that we are looking at in the returning customer model is LTV lift.

[00:09:19] LTV lift is the percentage of a cohort's first-time order revenue that comes back as returning revenue in each subsequent month. If a cohort spent $100,000 in their first month and month three LTV is lift is 2%, we expect $2,000 in returning revenue from that cohort three months later. LTV lift follows an exponential decay pattern, highest in the first month after acquisition, dropping sharply over the first several months and eventually flattening, flattening to a low, stable terminal rate that persists indefinitely.

[00:09:46] What we also know about the LTV impact is that it's realized pr- for most brands, large majority of the LTV impact is realized with- within the first 60 days or less, and it drops off very-- the decay drops off in a very steep curve following that time period. For any future month, returning revenue is the sum of all active cohort contributions.

[00:10:08] So each past cohort contributes its new revenue multiplied by the predicted LTV for that cohort's age. January 2026 returning revenue equals January run out plus December 2025 new revenue times the month one lift, plus November 2025 new revenue times the month two lift, plus October 2025 new revenue times the month three lift, and so on back through every active cohort of the customer file.

[00:10:32] The new revenue figures for each cohort from whatever plan is currently active at Statlas, which means the retention model is directly connected to the spending power model's output, and it's critical that these two models live in a connected ecosystem where one is informing the other. Otherwise, the amount of new customers that are coming in the future cohorts and how we are projecting out the new customer acquisition efficiency and volume isn't leading into a, a, an expectation of how that impacts the returning customer cohort and LTV lift over time.

[00:11:02] So the Profit Engineer Insiders system sees a cohort retention chart showing predicted LTV results over time, an actual versus model lift table comparing the data against predictions, and their regression results. This gives visibility into whether the brand's customer retention is improving, declining, or stable.

[00:11:19] A growing active customer file predicts revenue growth. A sh- a shrinking one sig-signals trouble, no matter how large the total email list looks. Many of those customers are lapsed, and they're not coming back. When something looks wrong as it relates to retention modeling is when the R squared, it sits at a, at a point that is at, at low confidence.

[00:11:41] So this is where your R squared might sit below point six percent. This model is gonna be flagged as underperforming and is not gonna give us high enough confidence to be able to make predictions against it. And so it should lead to more data cleaning, rerunning the model, focusing on tighter time frames, et cetera.

[00:11:58] Another, another time when a model the retention model might look off or that something's wrong is when the curve shape right-- shapes right and then levels off. So the data team adjusts the prediction scale, the pattern of decay is correct, the absolute values need recalibration. And so when the curve has a right shape and then levels off that's what might indicate that you don't have a strong enough retention model.

[00:12:21] And then the curve shape is wrong altogether. So if the retention behavior has fundamentally changed, the model's decaying to terminal rate, parameters are updated to reflect the new reality this is what might indicate that there's something in the buyer pers- in the buyer persona or behavior that has changed over time that that may require looking at a different timeframe relative to building, building the model.

[00:12:41] All things that we forecast across hundreds of brands, and so we're able to see and identify what a strong outcome of a retention model looks like and when there might be anomalies in the data that can allow us to build a more predictable a more predictable output. The fourth model is the event effect model.

[00:13:02] And you may if you're still with us here, you may be seeing that there are-- there's-- the third model has been skipped over which is the creative demand model. We are going to address this in more depth as-- when we walk through the creative strategy section of the canon and methodology. So we will get there.

[00:13:17] We're going to skip over the creative demand model as it connects deeply to the creative strategy process, and it'll be easiest to talk through there. So the fourth and final model as it relates to this section is the event effect model. So spending power model to retention model to the event effect model, which the whole goal of the event effect model is that we're getting daily precision through marketing events.

[00:13:37] The event effect model makes daily forecasting more accurate by learning how marketing events shift revenue patterns. It answers on a day with planned marketing event, how much should we expect new and returning revenue to deviate from the baseline? So there's two core signals. For every historical marketing event, the model measures two things: return ratio, how much does that event skewed revenue toward returning customers relative to a normal day?

[00:14:02] A ratio above baseline indicates the event activated the existing customer base more than usual, which is really com-common in loyalty campaigns, VIP drops, and seasonal events. The second signal the event effect model measures is AMER lift. How much more or less efficiently paid acqui-acquisition performed during the event compared to a baseline?

[00:14:20] A lift above one point oh means paid acquisition was more efficient, common in high intent sale periods where conversion rate spikes. So the event effect model helps us to understand the impact on your returning customer revenue ratio and then the AMER lift during specific moments over time. And this is informed by the data that sits at the base, at the foundation of the event effect model, which is over two years of marketing calendar event data as well as email and SMS behavior that is ingested into our system to form a qualitative base.

[00:14:54] And then those events are tagged by certain types. So the event the marketing event that was run in November of twenty twenty-four, is it a product launch? Is it a VIP drop? Is it a loyalty event? Is it a promotion? They're each tagged by the type of event that they are, and this forms a baseline for us to understand future events of that same type, so future promotional events, future product launch events, future VIP drop events, et cetera to understand the impact of those events on the return ratio and the AMER lift Based on the historical data set that exists, that helps us to get to daily precision against the forecast.

[00:15:35] So daily revenue distribution with events, event day spikes, surrounding days compressed, the month always reconciles. So we have the full month forecast based on the spending power model and the retention model, which is focused directly on those t-two core customer cohorts. The da- the event effect model helps us to flow the daily expectations as accurately as possible against the monthly expectation, so that we account for on those specific days with those specific event types, what is the impact going to be on the returning revenue ratio as well as the as well as the AMER lift or acquisition efficiency.

[00:16:13] The budget constraint guarantee. This is a really critical call-out. Event effects are applied as relative lifts, not absolute override. So if a month is planned at 30K spend at a 3.0 AMER, the aggregate daily outcomes across the month will always reconcile back to those targets. The model redistributes when revenue lands within the month and not the total amount.

[00:16:30] Event days run hotter, surrounding days compress proportionally. The month level constraint is never violated So how the profit engineer uses this inside of our system is that we build out the marketing calendar in Statlas for the coming month with planned expectations around the events that are going to be engaged in the coming, in the coming months.

[00:16:52] The profit engineer sees what the model predicts for each event type based on the historical data baseline, and the daily forecast adjusts op- automatically against that. Then from that baseline of the event effect model, we apply a day of week effect against this as well, which accounts for specific impact for days of the week that are stronger or weaker relative to their revenue contribution marg- contribution and behavior to the surrounding days.

[00:17:16] Every brand has a really distinct day of week effect related to the category that it sits in, the consumers that are oriented around the brand, and other factors. So the day of week effect plus the event effect model gets to a really strong expectation of what the daily expectations should be within a month for us to be able to achieve the monthly forecast so that we know as we're going through the month, are we pacing ahead or behind relative to where we should be at at this specific point in the month, accounting for the fluctuations in daily revenue, marketing calendar events, and daily events as well.

[00:17:52] So those core models, the spending power model, the retention model, and the event effect model form the foundation of how we forecast out future revenue based on the historical ingestion that sits against the baseline that we have in the account. These models are maintained by the CTC data team and tuned per brand.

[00:18:18] Each model has configuration controls that allow the data team to adjust for brand-specific behavior. The key principle is that models are tools, not oracles. They're built to be useful, not perfectly right. The value of the model is not in its point estimate, but in its ability to surface deviations quickly so the profit engineer can course-correct before the damage is done.

[00:18:37] We address this in depth in the forecasting section. So when a model's predictions deviate meaningfully from reality, the data team investigates. The question is not is the model broken, but what changed in the business that the model has not captured? Sometimes the answer is a retune. Sometimes the answer is that the business is in a generally new, generally new operating environment, and the model needs new data before it can be trusted against.

[00:19:01] So we have a standard cadence. We're on a quarterly basis. We're updating spending power models and retention mo- models. The event effect model updates continually as new events are added and then actualized. And then the creative demand model, which will be in the creative strategy section, is updated on a monthly basis.

[00:19:16] So these models form a foundation at a point in time, are updated continually based on their accuracy against the business obj- objective. And the operating principle is simple: when the map no longer matches the territory, update the map using this as the foundation of what informs the operational and executional plan we have on a day-to-day basis.

[00:19:38] That is the CTC canon of meth- of modeling following what we walked through on our methodology around forecasting in episode one, and we'll be coming with a handful of other deep dives into the canon of how we approach methodology across creative strategy, Facebook Ads, Google Ads, email strategy, and other areas on down the line as well.

[00:20:00] Thanks for hanging.