The question
How much might a repeat buyer spend over the next six or twelve months? Purchase history can help a team plan and prioritize, but a single customer’s future orders remain uncertain.
How the forecast works
One model estimates the probability of remaining active; another estimates monthly spend while active. Multiplying those estimates and adding the months produces a future-spend forecast. This example uses public wholesale transactions from Online Retail II, with values in pounds sterling.
Forecast a customer’s spend
Start with the prepared customer record. The result is an estimate, not a promise of future revenue. Changing inputs explores the model’s response; it does not establish the effect of an intervention.
Score a Customer
Customer 12352 is ready to score. Run the forecast first, then expand the purchase history to explore different inputs.
Adjust the customer’s purchase history
Customer 12352, a holdout record the model never trained on. Scores run against the live model API. Values are wholesale invoices in GBP.
How we judged usefulness
We compared the forecast with carrying prior spend forward and with a simple customer-segment average, using 1,295 customers held out of training. We checked forecasting error, customer ranking and predicted totals across groups.
Evaluation offline holdout
Lower forecasting error than a prior-spend baseline.
12-month WAPE: 68.8% vs 90.2% baseline; 1,295 held-out customers.
Customer-level forecasts remain noisy; use for aggregate planning and prioritization.
Evaluation data →What the result means
The model reduced forecasting error against both selected baselines. Ranking improved only modestly over prior spend, and individual estimates remain noisy. The evidence supports evaluating it for aggregate planning and prioritization, with care around decisions about any one customer.
For a different question—when customers leave and which segments retain better—see the customer retention case study. That example uses a separate sample dataset.
Results and BaselinesHoldout accuracy, customer rankings and comparison tables
Results
WAPE is weighted absolute percentage error on total spend, measured on customers the model never trained on: the sum of absolute errors divided by the sum of actual spend across the holdout customers. It is not the error of each customer. Customer-level spend on sparse wholesale buyers is noisy, so the measures that decide whether the model is useful are how well it ranks customers and how close predicted and actual spend sit by decile. The decile table and the baseline table below are that view. ROC-AUC is the survival model's range across the 12 forecast months. From metrics.json in the repo below. Spend is in GBP.
Where the Money Is: Predicted Against Actual by Decile, 12 Months
- Across all 1,295 holdout customers, predicted 12-month spend totals £1.49M against £1.55M actual, a 3.9% gap.
- The top two deciles by predicted value hold 70% of actual spend and are predicted within 1.3%. The small-buyer deciles are where the per-customer error lives; they hold little of the money.
| Decile by Predicted Value | Customers | Predicted, GBP | Actual, GBP | Gap |
|---|---|---|---|---|
| 1 | 130 | £7,061 | £17,842 | 60.4% |
| 2 | 129 | £13,302 | £24,782 | 46.3% |
| 3 | 130 | £23,786 | £39,706 | 40.1% |
| 4 | 129 | £35,284 | £29,881 | 18.1% |
| 5 | 130 | £47,504 | £38,300 | 24.0% |
| 6 | 129 | £61,409 | £60,261 | 1.9% |
| 7 | 129 | £85,437 | £105,554 | 19.1% |
| 8 | 130 | £141,622 | £148,034 | 4.3% |
| 9 | 129 | £250,293 | £251,611 | 0.5% |
| 10 | 130 | £828,030 | £838,734 | 1.3% |
Gap = absolute difference between the decile's predicted and actual totals, as a share of actual. From decile_table_12m in outputs/metrics.json.
The Model Against Two Naive Baselines, Same 1,295 Holdout Customers
- Repeat carries each customer's spend in the twelve months before the cutoff forward. Segment mean assigns the training-set mean for the customer's buyer-size tercile.
- The model improves over both selected baselines on every reported measure, but its absolute customer-level errors remain high. On 12-month WAPE it is 68.8% against 90.2% for repeat and 112.4% for the segment mean. On ranking, Spearman 0.612 against 0.600 and 0.573; the top predicted decile captures 53.9% of actual spend against 53.6% and 23.2%.
- An earlier Tweedie objective applied only to the spend component of the two-curve model returned 70.4% WAPE on this holdout, versus 68.8% for the serving model. That is different from the direct-forecast Tweedie experiment below.
| Measure, 12 Months | Model | Repeat Baseline | Segment-Mean Baseline |
|---|---|---|---|
| WAPE, 12 months (lower is better) | 68.8% | 90.2% | 112.4% |
| WAPE, 6 months (lower is better) | 82.8% | 125.1% | 122.2% |
| Spearman, predicted vs actual (higher is better) | 0.612 | 0.600 | 0.573 |
| Top-decile capture of actual spend (higher is better) | 53.9% | 53.6% | 23.2% |
From baselines, ranking, and spend_model_variants in outputs/metrics.json (buyer-value-radar f69559e). Both baselines use training rows only.
Alternative Model Comparison
A separate, exploratory check on 604 validation customers drawn from the training pool. Both approaches were trained and compared on the same customer split. These results are not from the 1,295-customer final holdout, and neither model was promoted as a result.
| Approach | 12-month WAPE ↓ | Top-decile spend capture ↑ |
|---|---|---|
| Two-curve model, retrained reference | 64.9% | 69.1% |
| Direct Tweedie forecast, challenger | 57.7% | 68.4% |
The direct Tweedie challenger had 7.1 percentage points lower WAPE on validation, but the paired-bootstrap 95% interval for the difference (Tweedie minus two-curve) runs from −14.4 to +4.7 points, so an improvement is not established. Top-decile capture was also slightly lower. This is a comparison of approaches, not a new production result. View the validation report →
Model and deployment details
How It Works
- Estimate ActivityProbability of remaining active
- Estimate SpendExpected spend while active
- Combine EstimatesActivity probability × monthly spend
- Sum Future ValueSix- and twelve-month forecasts
Code
The model, the training pipeline, and the scoring service: github.com/alphan-ml/buyer-value-radar
Where It Runs Data, deployment and checks
bvr-score-customer. Endpoint guijt78nlb.execute-api.us-east-1.amazonaws.com.giggit-buyer-value-radar-models) at cold start./health. GitHub Actions with an AWS role; no stored keys.