This is the third and final part of our ecommerce analytics series. In Part 1, we covered the five foundational metrics every merchant should track. In Part 2, we moved into cohort analysis, RFM segmentation, contribution margin, basket analysis, and discount effectiveness.
Now we're going further. The techniques in this guide — predictive LTV modeling, price elasticity, vendor concentration risk, channel attribution, and seasonal decomposition — are what separate merchants who react to their data from merchants who anticipate what's coming.
None of this requires a statistics degree. Each section explains the concept, gives you a practical approach you can use with real store data, and tells you what to do with the results.
Predictive LTV modeling: forecasting customer value before it happens
In Part 1, we defined customer lifetime value as the total revenue (or profit) a customer generates over their relationship with your store. That's backward-looking — it tells you what happened. Predictive LTV tells you what's likely to happen next.
Why prediction matters
If you know that customers acquired through Instagram ads have a predicted LTV of $320, while customers from Google Shopping have a predicted LTV of $180, you can bid more aggressively on Instagram and set tighter return targets for Google. Without prediction, you're treating all new customers as if they're equally valuable.
Predictive LTV also changes retention spending. If an at-risk customer has a predicted remaining LTV of $400, spending $50 on a win-back campaign makes sense. If their predicted remaining value is $20, it doesn't.
A practical approach merchants can use
You don't need a machine learning pipeline. A simple historical model works well for most merchants:
Step 1: Calculate average revenue per customer per month, by cohort. Use your cohort analysis from Part 2. For each monthly cohort, divide total revenue by the number of customers in that cohort for each month of their lifespan.
Step 2: Build a revenue decay curve. Plot average monthly revenue per customer over time. Most ecommerce businesses see a pattern: high spending in Month 0 (the first purchase), a sharp drop in Month 1, then a gradual decline that flattens out around Month 6-12. The customers who survive past Month 3-4 tend to stabilize.
Step 3: Extrapolate the curve forward. If your Month 6 average revenue per surviving customer is $28 and it's been declining by roughly 5% per month, project that forward: Month 7 is $26.60, Month 8 is $25.27, and so on. Sum the projected months to get the predicted remaining LTV.
Step 4: Adjust by segment. Don't use one model for all customers. Run separate curves for your RFM segments. Champions have a flatter decay curve (they keep spending). New Customers have a steeper one (most don't come back). At-Risk customers fall somewhere in between.
Making it actionable
- Set acquisition cost ceilings per channel based on predicted LTV
- Allocate retention budget proportionally to predicted remaining value
- Identify which customer segments are trending toward higher or lower LTV over time, and investigate why
- Flag customers whose actual spending deviates significantly from prediction — both positive (opportunity to upgrade) and negative (early churn signal)
Price elasticity: understanding how price changes affect demand
Price elasticity measures how sensitive your customers are to price changes. If you raise the price of a product by 10% and unit sales drop by 15%, demand is elastic — customers are price-sensitive. If unit sales drop by only 3%, demand is inelastic — customers are relatively indifferent to the price change.
Why this matters for merchants
Most merchants set prices based on cost-plus (COGS + desired margin) or competitive benchmarking (match or undercut competitors). Neither approach tells you whether you're leaving money on the table.
A product with inelastic demand can sustain a higher price without losing meaningful volume. A product with elastic demand needs competitive pricing or it bleeds sales. Without measuring elasticity, you're guessing at where that line is.
How to test price sensitivity
A/B price testing: If your platform supports it, show different prices to different customer segments and measure conversion rate and revenue per visitor. This is the cleanest test but requires enough traffic to reach statistical significance. A product with 50 views per week won't give you reliable results for months.
Sequential testing: Change the price for a defined period (two to four weeks), then change it back. Compare unit sales volume, revenue, and margin between the two periods. Control for seasonality and promotions — a price test during a holiday week won't give you clean data.
Natural experiment analysis: Look for past price changes in your data (intentional or due to supplier cost changes) and measure the before-and-after impact on volume. This is retrospective, but if you have the data, it costs nothing to analyze.
Measuring the impact
The price elasticity formula is straightforward:
Elasticity = % Change in Quantity Sold / % Change in Price
If you raised the price by 10% and quantity dropped by 20%, elasticity is -2.0. The negative sign is standard — price up, quantity down.
Rules of thumb for interpretation:
| Elasticity | Meaning | Action |
|---|---|---|
| 0 to -0.5 | Highly inelastic | Room to raise price |
| -0.5 to -1.0 | Moderately inelastic | Price increase may work; test carefully |
| -1.0 | Unit elastic | Revenue is maximized at current price |
| -1.0 to -2.0 | Elastic | Price decrease may increase total revenue |
| Below -2.0 | Highly elastic | Customers are very price-sensitive |
Practical considerations
- Elasticity varies by product category. Commodity products (phone chargers, basic tees) tend to be elastic. Differentiated or niche products tend to be inelastic.
- Elasticity can change over time. A product that was inelastic when you were the only seller becomes elastic when competitors enter.
- Always measure revenue impact, not just volume. A 10% price increase that causes a 5% volume drop increases total revenue by about 4.5%.
- Test on your mid-tier products first — they have enough volume to generate reliable data and enough margin to absorb a test that goes wrong.
Vendor concentration risk: how exposed are you?
If 60% of your revenue comes from products supplied by a single vendor, you have a concentration problem. If that vendor raises prices, has a supply disruption, or discontinues a product line, your business takes a disproportionate hit.
The Herfindahl Index, simplified
The Herfindahl-Hirschman Index (HHI) is a standard measure of concentration. For merchants, it applies to vendor (supplier) concentration. The calculation is simple:
- Calculate each vendor's share of your total COGS (or revenue — COGS is more relevant since it measures supply dependence)
- Square each share
- Sum the squared shares
Example with four vendors:
| Vendor | % of COGS | Squared |
|---|---|---|
| Vendor A | 55% | 0.3025 |
| Vendor B | 25% | 0.0625 |
| Vendor C | 12% | 0.0144 |
| Vendor D | 8% | 0.0064 |
| HHI | 0.3858 |
An HHI above 0.25 indicates high concentration. Above 0.40 is dangerously concentrated. The example above (0.3858) is in the danger zone, driven almost entirely by the 55% dependence on Vendor A.
For comparison, a perfectly diversified portfolio across four equal vendors would have an HHI of 0.25 (each at 25%, 4 x 0.0625 = 0.25). Across ten equal vendors, it drops to 0.10.
Diversification strategies
Dual-source critical products. For any product that represents more than 15% of your revenue, establish a secondary supplier. You don't need to split volume 50/50 — even a 70/30 split gives you fallback capacity.
Monitor vendor concentration quarterly. HHI can creep up as popular products from one vendor grow faster than the rest of your catalog. Track it over time so you catch the drift before it becomes a crisis.
Negotiate from knowledge. When you know your HHI and can show a vendor they represent 55% of your supply chain, it changes the negotiation dynamic. You can either negotiate better terms (you're their biggest customer by volume) or start diversifying (you're too dependent on them).
Build private-label alternatives. For commoditized product categories, developing your own brand version reduces vendor dependence and typically improves margins. This is a longer-term play but worth considering for your top-selling categories.
Channel attribution: knowing which channels actually drive sales
If a customer sees your Facebook ad, clicks a Google search result a week later, and then converts through a direct visit, which channel gets credit? The answer determines where you spend your marketing budget.
The problem with last-click
Most analytics platforms default to last-click attribution: the last touchpoint before conversion gets 100% of the credit. In the example above, "direct visit" gets the credit. Facebook and Google get nothing.
This systematically undervalues awareness channels (social, display, content) and overvalues capture channels (branded search, direct, retargeting). You keep cutting awareness spend because it "doesn't convert," and then wonder why your direct and branded search traffic declines months later.
Multi-touch models
Several models distribute credit more fairly:
Linear: Every touchpoint gets equal credit. Simple but doesn't account for the fact that some touchpoints matter more than others.
Time decay: Recent touchpoints get more credit than earlier ones. Intuitive — the closer to purchase, the more influence.
Position-based (U-shaped): First touch (awareness) and last touch (conversion) each get 40%. Everything in between splits 20%. A good default for most merchants.
Data-driven: Uses your actual conversion data to calculate each touchpoint's contribution statistically. Most accurate but requires significant data volume.
Incrementality testing
Attribution models estimate channel value. Incrementality testing measures it directly: pause a channel for a test group for 2-4 weeks and compare conversions against a control group still seeing ads. The difference is that channel's true incremental contribution.
If pausing Facebook ads causes a 12% conversion drop, Facebook drives 12% incremental lift. If conversions barely change, the channel was taking credit for sales that would have happened anyway. Run these tests quarterly on your highest-spend channels.
Practical steps
- Move beyond last-click as soon as possible — it distorts every budget decision
- Start with position-based attribution as a reasonable default
- Run incrementality tests on any channel consuming more than 20% of your budget
- Track channel performance on a contribution-margin basis, not just revenue. A channel driving high revenue through discount-dependent customers may have negative margins
Seasonal decomposition: separating signal from noise
Every ecommerce business has seasonality — revenue patterns that repeat on a weekly, monthly, or annual cycle. Black Friday spikes. Summer lulls. End-of-month paycheck bumps. The problem is that seasonality makes it hard to tell whether your business is actually growing, shrinking, or just riding the calendar.
What decomposition does
Seasonal decomposition breaks your revenue (or any time-series metric) into three components:
Trend: The underlying long-term direction. Is the business growing at 3% per month, flat, or declining? Trend strips out the noise and shows the trajectory.
Seasonality: The repeating calendar pattern. Revenue goes up 25% every November and down 15% every January. This component is predictable and cyclical.
Residual (noise): Everything left over — unexpected spikes, one-time events, random variation. A surprise viral post. A supply-chain disruption. A competitor going out of business.
When you separate these three components, you can answer questions that raw data obscures:
- "Is the business growing, or did we just have a good seasonal month?"
- "Was this month's dip a real problem or just normal seasonality?"
- "If I remove seasonal effects, what's my true month-over-month growth rate?"
A practical approach
Step 1: Collect at least 12-18 months of monthly revenue data. You need at least one full cycle to identify seasonal patterns. Weekly data works too, but monthly is easier to work with and less noisy.
Step 2: Calculate moving averages. A 12-month centered moving average smooths out seasonality and gives you the trend component. For each month, average the 6 months before and after it (plus the month itself). The resulting curve is your trend line.
Step 3: Calculate seasonal indices. Divide each month's actual revenue by the trend value for that month. This gives you a ratio. Average the ratios for each calendar month across all available data. The result is a seasonal index — January might be 0.82 (18% below trend), November might be 1.35 (35% above trend).
Step 4: Deseasonalize your data. Divide each month's actual revenue by its seasonal index. The result is "seasonally adjusted" revenue — what you would have earned if there were no seasonal effects. Now you can compare January to November on an equal footing.
Using decomposed data for forecasting
Once you have trend and seasonal components, forecasting is straightforward: project the trend forward, then multiply by each month's seasonal index. Your November forecast = projected trend value x 1.35. Use the residual range (typically +/- 8%) as your confidence interval.
This won't compete with sophisticated forecasting models, but it's dramatically better than "last month plus 5%" or "same month last year plus growth."
Practical applications
- Inventory planning: If November demand is 35% above trend, start building stock in September
- Cash flow management: A January dip of 18% is not a crisis if it happens every year — plan reserves accordingly
- Marketing budget: Concentrate ad spend in months where seasonal demand is already rising
- Performance evaluation: Judge results against seasonally adjusted expectations, not raw comparisons
Bringing it all together
These five advanced techniques connect to each other and to the intermediate techniques from Part 2:
- Predictive LTV builds on the cohort analysis and RFM segmentation you already set up
- Price elasticity feeds into your contribution margin analysis — you can model the margin impact of price changes before making them
- Vendor concentration affects your cost structure, which flows through to contribution margins
- Channel attribution tells you where to find more of your highest-LTV customer segments
- Seasonal decomposition makes every other metric more accurate by removing calendar noise
Start with the technique that addresses your biggest blind spot. Heavy marketing spend with unclear returns? Start with attribution. Prices set by gut feel? Start with elasticity. Revenue chart looks like a roller coaster? Start with seasonal decomposition.
Spark can run each of these analyses on your connected store data — LTV projections, seasonal patterns, vendor concentration scores, and margin impact analysis — without spreadsheets or data engineering.
This wraps up the three-part ecommerce analytics series. Part 1 gave you the foundation. Part 2 gave you the intermediate toolkit. This guide gave you the advanced lens. The common thread: analytics isn't about more data or fancier tools. It's about asking better questions and acting on the answers.
Advanced analytics without the complexity
Spark runs predictive models, seasonal analysis, and margin calculations on your store data — no spreadsheets required.