Why Predictor Preparation Belongs in a Business Report
A predictive model does not always use source fields exactly as they first appear in a data file. A dollar amount might be rescaled, several low-volume categories might be combined, or two recorded fields might be used to calculate a rate. These steps determine the inputs the model receives. If a report omits them, a reader may misunderstand what a reported factor means or assume the model used a source value in its original form.
A useful explanation is specific enough to be checked and understood without requiring the reader to interpret code. State what was changed, how the change was made, why it was made, and what the model receives afterward. Distinguish preparation of an input from a change to the underlying observed record: converting dollars to thousands changes the numerical representation, not the amount owed.
Explain the Operation, Not Just Its Name
Terms such as “normalized,” “cleaned,” or “engineered” may be familiar to analysts but do not tell a decision-maker what happened. Replace labels with a plain-language description and, when material, the exact rule. For example, “annual premiums were divided by 1,000, so the model input is measured in thousands of dollars” is more informative than “premium was normalized.”
A concise explanation usually addresses four questions:
Name the recorded field or fields, their units, and any relevant observation period.
State the formula, conversion factor, cut points, or category mapping when these affect interpretation.
Describe the resulting value or category, including units and any new indicator inputs.
Give the practical reason, such as making units easier to read or combining categories with few observations, without implying more than the operation establishes.
The purpose should be accurate and proportionate. A category grouping may make inputs more manageable, but it does not prove that the grouped categories have identical outcomes. A rate may make exposure levels easier to compare, but it does not by itself establish why one record has a higher rate. Describing predictor preparation is not a claim that the operation removes bias or improves predictions.
Scale Changes and Numeric Transformations
For numeric inputs, report the original unit, the transformation, and the resulting unit. If a field is divided by a constant, say what that constant is and why the new scale is being used. If values are centered or standardized, state the reference mean and scale when they are needed to interpret the transformed value. Do not use “standardized” as a substitute for describing the actual operation.
Worked Example: Reporting a Rescaled Claim Amount
A hypothetical claims team records claim amounts in dollars. Before fitting a model, the analyst divides each amount by 1,000. One claim is $2,750. Explain the preparation in business language and calculate the input supplied for that claim.
Method. Identify the source unit, apply the stated conversion, and label the resulting unit. The transformation is \(z=x/1{,}000\), where \(x\) is the recorded amount in dollars and \(z\) is the model input in thousands of dollars.
Calculation. The recorded amount is \(x=\$2{,}750\). Dividing by 1,000 gives \(z=2{,}750/1{,}000=2.75\). Thus, the model receives 2.75 in units of thousands of dollars; it does not receive 2.75 dollars.
Report wording. “Claim amounts were converted from dollars to thousands of dollars by dividing each recorded amount by 1,000. For example, a recorded amount of $2,750 enters the model as 2.75 thousand dollars. This changes the input scale, not the claim amount.”
Check and common mistake. Since \(2.75\times \$1{,}000=\$2{,}750\), the conversion is consistent. Calling 2.75 “$2.75” would be wrong: it drops the unit conversion and understates the amount by a factor of 1,000.
Centering and standardizing require more context than a simple unit conversion. For example, if \(z=(x-\mu)/s\), the report should identify \(\mu\) as the centering value and \(s\) as the scale used, if readers need to interpret \(z\). A transformed value is a distance in the chosen scale, not a new observed measurement in the original unit.
Why an Affine Scale Change Can Preserve a Linear Predictor
A short algebraic justification helps distinguish a change in input scale from a change in the information recorded. Consider a model whose prediction function uses a linear predictor \(\eta=\beta_0+\beta_1x\). Suppose the same input is represented as \(z=(x-\mu)/s\), where \(s\ne 0\). Then \(x=\mu+sz\), and substitution gives:
Thus, with coefficients re-expressed as \(\beta_0^*=\beta_0+\beta_1\mu\) and \(\beta_1^*=\beta_1s\), the same linear predictor can be written \(\eta=\beta_0^*+\beta_1^*z\). This identity applies to an affine rescaling in a linear predictor when the coefficients are adjusted accordingly; it does not say that the coefficient values stay the same. It also does not establish that every modeling procedure or model constraint is unaffected. The practical reporting point is narrower: identify the new units and avoid interpreting a coefficient as though it still referred to a one-unit change in the original field.
Category Groupings and Encoded Inputs
A categorical field can be prepared by combining labels or by representing categories with indicator inputs. Explain both the mapping and the model-facing result. If “north district” and “central district” are combined as “urban service area,” state that mapping explicitly and disclose any other labels left separate. If the model uses indicator inputs, a reader generally needs to know which category is the reference and what a value of one means.
| Preparation detail | Plain-language report question |
|---|---|
| Grouping | Which original labels were combined, and what is the new label? |
| Unchanged categories | Which labels remain separate? |
| Reference category | Which category is represented by the baseline rather than a separate indicator? |
| Unseen or missing labels | What happens if an input label is missing or was not among the categories used in preparation? |
Worked Example: Explaining a Category Grouping and Its Indicators
A hypothetical service-request model uses four recorded regions. The preparation rule combines “Ridge” and “Valley” into “Remote”; “Harbor” and “Plains” remain separate. The model uses “Harbor” as its reference category and creates indicator inputs for “Plains” and “Remote.” A record from Ridge is processed. Describe the mapping and calculate its indicator values.
Method. Apply the mapping first, then assign one to the indicator matching the resulting category and zero to the other listed indicators. The reference category has no separate indicator in this coding.
Mapping. Ridge maps to Remote. The record is therefore in the Remote category after preparation.
Indicator calculation. The record is not Plains, so the Plains indicator is 0. It is Remote, so the Remote indicator is 1. Its model-facing category inputs are therefore Plains \(=0\) and Remote \(=1\). A Harbor record would have Plains \(=0\) and Remote \(=0\), representing the reference category.
Report wording. “Ridge and Valley records were combined as Remote; Harbor and Plains remained separate. Harbor is the reference category. The model receives indicators for Plains and Remote, so a Ridge record is represented as Remote = 1 and Plains = 0.”
Check and common mistake. The two indicators identify one prepared category without assigning Ridge its own separate input. Saying that Ridge has an indicator of one would be incorrect: it maps to Remote before indicators are assigned. The report should also state how missing or previously unseen region labels are handled if they can occur in the intended use.
Derived Features: Show the Recipe and the Meaning
A derived feature is calculated from one or more recorded fields. Its explanation should make the calculation reproducible and its meaning clear. Name the inputs, the units, the time period, and any rule for zero or missing denominators. If the new feature is a ratio, distinguish the numerator from the denominator. If it is an indicator, state the threshold and whether equality is included.
Worked Example: Explaining a Service-Use Rate
A hypothetical equipment-support team records, for each account, 18 service contacts during a six-month observation window and 12 covered equipment units during that same window. A derived feature is contacts per covered unit. Calculate the value and write a report explanation that does not imply a causal conclusion.
Method. Divide the number of contacts in the stated window by the number of covered units. The resulting unit is contacts per covered unit over six months.
Calculation. The numerator is 18 contacts and the denominator is 12 units. Therefore, \(18/12=1.5\) contacts per covered unit over six months. As a check, \(1.5\times12=18\) contacts.
Report wording. “We calculated six-month service contacts per covered unit by dividing each account’s contact count during the six-month window by its covered-unit count for that window. An account with 18 contacts and 12 units has a value of 1.5 contacts per unit. This summarizes recorded use per unit; it does not establish why the account required service.”
Check and common mistake. Both fields refer to the same six-month window, so the ratio compares consistent periods. Using annual contacts with a six-month unit count would not represent the stated feature. The report should also specify how records with zero covered units are treated, because division by zero is undefined; no such rule can be inferred from the example data.
Derived inputs can make a relationship easier to express, but they can also hide detail. A rate may conceal different numerator and denominator values: 2 contacts per unit could arise from 2 contacts and 1 unit or from 20 contacts and 10 units. When that distinction matters to the intended decision, report the feature alongside its component fields or explain the limitation.
Materiality, Reproducibility, and Responsible Communication
Not every code-level operation belongs in a decision-focused report. A transformation is material when omitting it could change how a reader interprets a model input, the population to which the rule applies, or the operational use of the model. Disclose the rules that define categories, alter units, create thresholds, or combine source fields. A brief technical appendix can hold implementation details, while the main explanation states the business meaning.
Before finalizing the description, check that it answers these practical questions:
- Can a reader tell which source fields contribute to each prepared input?
- Are units, cut points, category mappings, and reference categories stated where needed?
- Could the rule be applied consistently to a new record, including missing or unfamiliar values?
- Does the wording distinguish a mathematical operation from a claim about causes, fairness, or improved performance?
- Could a grouping hide meaningful differences between the original categories? If so, is that limitation stated?
A useful standard is reproducibility: another analyst with the same source fields and stated rules should be able to determine what inputs a record would receive. This does not require presenting code in the main report. It does require avoiding vague descriptions that leave important decisions, such as category boundaries or denominator handling, unstated.
Check Your Understanding
For each situation, identify the model-facing value or representation and the key detail a business report should disclose.
- A hypothetical model uses repair cost in thousands of dollars. A recorded cost is $6,400, and the preparation divides dollars by 1,000. What value enters the model, with units? (Answer: 6.4 thousand dollars.)
- A “channel” field groups Web and Mobile as Digital; Branch remains separate. Branch is the reference category, and indicators are created for Digital. How is a Mobile record represented? (Answer: Mobile maps to Digital, so the Digital indicator is 1; Branch is the reference.)
- An account has 24 visits during a three-month window and 8 active locations during the same window. A feature is visits per active location. Calculate it and state its unit. (Answer: \(24/8=3\) visits per active location over three months.)
- A numeric input is centered and scaled as \(z=(x-50)/10\). For a source value of 70, calculate the prepared value and name the interpretation. (Answer: \(z=(70-50)/10=2\), two scale units above the centering value.)
- A hypothetical feature is an indicator equal to 1 when a balance is at least $500 and 0 otherwise. A balance is exactly $500. What is the feature value, and what wording must the report make clear? (Answer: 1; the threshold is inclusive, so balances equal to $500 are included.)