Predictive Modeling

Know how your market responds before you change a price.

Measure price sensitivity, forecast demand and explain discounting with models fitted to your own commercial data, each checked before you rely on it.

Decision Optimization / Predictive Models / Price Elasticity on Invoice lines
Question
Can a distributor raise fittings prices without losing too much volume? Illustrative data from one fictional business.
Evidence
The 1/2" coupling loses about 4.1% of volume for each 1% price rise; the 1" ball valve loses about 0.74%, so it tolerates a rise.
Limit
The 4" cleanout has only 24 price points and a range that crosses zero, so it is not significant. Treat its figure as a rough guide.
01 · Measure the response

Build the model behind the answer.

The question is how fittings volume reacts to price, so the model is a price elasticity on invoice lines: pick the type, map the columns, and it says how it will check itself.

Model types

Ten questions, each with its own model.

Each type names the question it answers and how it learns: fixed rules, a statistical fit, or machine learning that checks itself on data it did not see.

Columns

Map your columns, then add drivers.

Each type asks for the columns it needs. Extra columns can change the model, split it into one per value, or ride along into the results.

How it works: model types and their checks
Approaches
RFM Segmentation uses fixed rules and trains nothing. Price, cross-price and promotion models fit a formula every run inside the database. Forecasts, clustering and drivers train with machine learning in the background.
Profile first
Before a run, a data profile shows what the model will see, such as how many groups clear the minimum number of price points, and recommends preparation like capping extreme prices.
Methods
Where a type has several methods, such as forecasting, you choose which to try, the measure to judge them by, and whether to keep the best one or an average.
Plain explanations
Every type explains the models it uses and how it is checked before you run it, in the same words the results and the report use.
02 · Test the price move

Try the increase before it reaches a customer.

Score candidate prices for the same fittings against the fitted elasticities, then read where the data supports a higher price and where it points the other way.

Decision Optimization / Predictive Models / Price Elasticity on Invoice lines

The Score tab of the same price elasticity model. Rows with a unit cost show profit and margin; the last row, with no product, moves every product with a clear price signal by 2%. Illustrative data.

  • One product or all of themName a product and a price, or leave the product empty to move every product with a clear signal by the same percentage.
  • Warnings where the data is thinA price the product never sold at, or an elasticity without a clear signal, is flagged beside the result.
  • Many rows at onceAdd rows one by one or paste them from a spreadsheet, up to 5,000 at a time.
Price optimization

Find the profit-maximising price, where the data supports one.

A price optimization model reads each product’s elasticity and cost and gives the price that maximises profit, kept close to the prices you have actually charged.

What the price models assume

Read the result with its limits.

  • Measured from your historyElasticities come from how volume moved when your own prices moved. Where prices barely changed, the result is flagged rather than guessed.
  • Other things held steadyThe models assume competitor prices and promotions stayed constant, and they are least reliable far outside the prices you have charged.
  • Profit, not your strategyThe optimal price ignores volume targets, price points and effects between products. Decide those in Decision Optimization.
Decision Optimization
How it works: scoring and the optimal price
Scoring
Score uses the version in use. Expected units scale by the price change raised to the product’s elasticity, over a period as long as the data the model learned from. Profit and margin need a unit cost, typed on the row or carried by the model.
Optimal price
The optimum is cost × e ÷ (e + 1), given only where the elasticity is significant and clearly elastic, and kept within the observed price range widened by an allowed stretch (10% unless you change it).
No price is changed
Scores and optimal prices are results to review. Nothing is written to a price list until someone acts on it in Price Management.
Not a constraint solver
There are no floors, price bands or deal rules here. Those belong to Decision Optimization.
03 · Find where it leaks

Check where discounting would undo the increase.

A higher price only holds if discounts stay in range. Ridgeway Plumbing buys the most price-sensitive product, gets more discount than its group normally does, and has not ordered recently.

Discount drivers

Find who gets discounted, and where it runs too deep.

The model learns which customer, product and order attributes explain discounting, as readable rules, and sizes the revenue above each group’s normal range.

Customer segments

Group customers by how they buy.

Score every customer on how recently, how often and how much they buy, or let clustering find the groups in the data.

How it works: drivers and segments
Drivers explain, they do not prove
Discount and transaction drivers come from a readable decision tree, checked on customers or transactions it did not see. They show which attributes go with deeper discounts, not that the attribute caused them.
Revenue at stake
Each group gets a tail threshold, the point above which a discount counts as too deep for that group. Revenue at stake is the excess discount times the revenue at list price.
Segments
RFM scores each customer from 1 to 5 against the rest of the base and names the segment with fixed rules. Customer and transaction clustering find groups from the data instead, and report how well separated they are.
Also available
Cross-price elasticity shows which products take sales from, or support, each other. Promo uplift measures whether promotions bought extra volume and what they cost in margin.
04 · Plan and keep current

Forecast the quarter, and keep the model current.

Forecast demand for the fittings into the next quarter, and let the elasticity model retrain as new invoices arrive, so the next price review starts from current data.

Demand forecast

Forecast the quarter for the same fittings.

Each product is forecast from its own history, with a likely range, and the run writes a report that explains the result in charts and plain language.

Versions

Retrain on new data, and keep the better model.

The elasticity model retrains when its source refreshes. A new version replaces the one in use only when it is not less accurate.

How it works: forecasts, versions and reports
Forecasts
Per series, the model tries smoothing, Holt-Winters, ARIMA, Prophet and gradient boosting, plus one model trained across all series, and keeps the best on rolling back-tests. The likely range is calibrated on those back-test errors.
Versions
The ten most recent versions are kept with their results. A run you start becomes the version in use; a retrain started by a refresh only does if it is not less accurate. Switch back to an earlier version at any time.
Methods
Where a type can fit several curves or methods, each run scores them on the measure you choose and keeps the best, or an average.
Reports, Copilot and export
Generate a written report of each run, ask Copilot to create, run, compare or score a model, and export results to CSV. Lineage shows which models depend on a dataset or view.
05 · What’s included

What’s included.

What Predictive Modeling gives you, what every subscription adds, and what to add later.

In Predictive Modeling

  • Ten model typesForecasts, price response, drivers and segments, each with its own checks.
  • Score and optimal priceTry prices against a fitted model before you change one.
  • Versions and reportsQuality per version, retraining on refresh and a written report of each run.

In every subscription

Add when you need it

Solutions it supports: Revenue & Margin Planning, Manage Quotes & Deal Pricing, Price Realization & Pricing Leakage.

FAQ

Common questions

What does Predictive Modeling do?

It fits models to your own commercial data to answer specific questions: how volume reacts to price, how much you will sell next period, which customers buy alike, and what explains who gets discounted. There are ten model types today, each with its own checks, and the results feed planning, pricing and quoting decisions.

What data do I need?

It depends on the model. Price elasticity needs price and quantity on each sale, a demand forecast needs a date and a value, customer segments need a customer, a date and spend, and discount drivers need the discount and the revenue it applied to. Add a cost column for profit and margin. The data comes from datasets and views in Data & Context Studio.

How do I know whether to trust a result?

Every model type declares how it is checked, and the results show it. Elasticities carry a p-value and a 95% range, groups with too few price points are flagged, forecasts are back-tested on held-back periods, and each version records its quality figure. A result the data cannot support is left empty rather than guessed.

Can it recommend prices?

Partly. A Price Optimization model gives the profit-maximising price per product from its measured elasticity, only where the data supports one and kept close to prices you have actually charged. It does not apply floors, price bands or approvals. Choosing prices within those limits is Decision Optimization, and changing them is Price Management.

Does it show what causes a change?

No. Elasticities and drivers measure how things moved together in your history, which is useful evidence but not proof of cause. The price models assume competitor prices and promotions held steady, and each report lists the assumptions and limits of its run so you can judge how far to rely on it.

How do models stay current?

Run a model again, update it with the latest data, or turn on retraining so it trains a new version each time its source refreshes. A retrained version only replaces the one in use when it is not less accurate. Recent versions are kept, so you can compare two and switch back.

Is our data used to train models?

Your data is used to provide, support, secure and improve the Service. Separately, the agreement permits Revomo to use aggregated and de-identified data, including to train and improve models, only in a form that cannot identify you, your users or your customers. Third-party models are used for inference through Amazon Bedrock, a listed subprocessor.

Do predictions change prices on their own?

No. A model’s results are an input to a decision, not an action. Configured constraints apply to recommendations, exceptions route for review, and actions proceed through approval workflows, each with lineage and an audit trail. An action that needs a person’s approval still needs it when an agent proposes it.

See how your customers respond to price.

Walk through these models on sample data, and talk through which questions your own data can answer.

Plate B2 · Series 2026