Held out model comparison, 84,000 records, one operator

Data science research poster example

A model comparison poster is read as a table first, because the ranking is the finding. This board puts five scorers in one table, the rules list currently in use and four models, all measured on the same held out set of 25,200 subscribers, and leads with precision in the top decile rather than with accuracy.

Create a poster with OneCraftA0 Landscape, printed at 1189 by 841 millimetres

The whole board

The poster at full size, exactly as it prints. Every number, citation and caption on it was written for this example, so the layout is being judged on real content.

Four churn models against the rules list they would replace: 84,000 subscriber records, one test set

D. Marchetti ¹, S. Bhandari ¹

1 Applied Data Science Group, Thornbury Institute of Technology

Current practice

Retention teams at the partner operator score churn risk with a hand written rules list: two late payments, a support contact in the last month, a plan older than three years. The list is readable and it is cheap, and nobody has measured it against an alternative on the same records. Scored over the same held out test set it reaches an AUC of 0.58 and finds 479 of the 3,024 churners in the top decile, a precision of 0.19. That is the bar every model on this board has to clear. References. 1. Oyelaran, F. (2025) Held out evaluation for retention models. Journal of Applied Analytics, 14(1), 22 to 40. 2. Prasetyo, R. and Kim, L. (2024) Precision at k against AUC in campaign targeting. Data Practice Review, 9(2), 88 to 104. 3. Marchetti, D. (2024) Rules lists as a baseline nobody scores. Thornbury Working Paper 2024-11.

PROPOSED PROTOCOL

Records. 84,000 subscriber records from one operator, 24 months of billing, usage and support history, churn labelled at 30 days after contract end.

Split. A 70 to 30 split by subscriber, never by row, giving 58,800 training records and 25,200 test records. The test set holds 3,024 churners, 12.0%.

Features. 38 features: tenure, plan age, billing amount and variance, late payments, data and call volumes, support contacts and handset age.

Tuning. Five fold cross validation inside the training set only. No test record was seen during tuning, feature selection or threshold choice.

Measures. Area under the ROC curve for ranking quality, and precision in the top decile of 2,520 subscribers, which is the size of one campaign.

Five scorers, one test set

Scorer	AUC	Top decile precision
Rules list in use today	0.58	0.19 (479 of 2,520)
Logistic regression	0.71	0.34 (857 of 2,520)
Random forest	0.78	0.41 (1,033 of 2,520)
Gradient boosting	0.81	0.46 (1,159 of 2,520)
Neural network	0.80	0.44 (1,109 of 2,520)
Held out test set of 25,200 subscribers containing 3,024 churners.

12 points

more top decile precision than logistic regression

Gradient boosting 0.46 against 0.34, on the same 2,520 subscriber campaign

AUC

Precision

Figure 1. Left bar AUC, right bar top decile precision, for the rules list, logistic, forest, boosted and neural scorers.

Source: one operator, 24 months of records, 2026.

Conclusions

Gradient boosting ranks best on both measures and the neural network is close enough that the choice between them should rest on cost to run rather than on these numbers. The gap that matters is not between the two leaders, it is between either of them and the rules list: 1,159 churners found against 479, in a campaign of the same size and cost. Recall in that decile rises from 15.8% to 38.3% of all churners in the test set.

Gradient boosting AUC 0.81, top decile precision 0.46

The rules list in use scores 0.58 and 0.19 on the same test set

One campaign of 2,520 finds 1,159 churners against 479

Neural network within 0.02 AUC of the leader on 38 features

Block by block

What each block on the board is for, in the order a reader walks it.

Title band and two authors: The title names the comparison and the record count, and says the rules list is the thing being replaced, which is the framing the rest of the board depends on.
Current practice: What the retention team scores churn with today, why nobody had measured it, and its scores on the same held out set: AUC 0.58 and 479 of the 3,024 churners in the top decile. Three numbered citations close the block.
Proposed protocol: Five labelled entries: the 84,000 records and the 30 day churn label, the 70 to 30 split by subscriber rather than by row, the 38 features, five fold cross validation inside the training set only, and the two measures reported.
Five scorers, one test set: A three column table listing every scorer with its AUC and its top decile precision, the raw count printed beside each rate, over a test set of 25,200 subscribers holding 3,024 churners.
Headline number: The stat block carries 12 points, the gap between gradient boosting and logistic regression in top decile precision, with both values and the campaign size underneath.
Figure 1, both measures by scorer: Two bars per scorer, AUC and top decile precision, on a shared zero to one scale. This figure block has its legend switched off, so the caption names which bar is which.
Conclusions and four checked points: Why the choice between the two leaders is a cost question, why the gap that matters is against the rules list, and a four item list giving the leader scores, the incumbent scores, the campaign counts and the neural network result.

What makes this board work

The incumbent is scored too

Four models against each other is an academic exercise. Four models against the rules list a team uses today, on the same test set, is a decision. Adding that fifth row turned a 12 point gap between two models into a 27 point gap that a retention manager can act on.

Precision at a real campaign size

AUC ranks well and means little to anyone buying a campaign. The board reports precision in a top decile of 2,520 subscribers, which is one campaign, and prints the raw counts: 1,159 churners found against 479 for the same cost.

The layout forces the comparison

This design is built around a comparison table with a current practice column on the left and a proposed protocol column on the right. Putting the rules list in one and the modelling protocol in the other uses the layout for what it was drawn to do.

Questions people ask

What goes on a data science research poster?

The dataset and its size, how it was split, the features, the tuning procedure and where it ran, the measures and why those measures, and the results as a table. A model comparison poster should also score whatever the team uses today, because that is the baseline a reader is really deciding against.

Should I report AUC or precision at k?

Both, and label which one the decision rests on. AUC describes ranking across the whole population, while precision in the top decile describes the one campaign anybody will actually run. This board gives both for five scorers and leads with the decile, because 2,520 subscribers is the size of the campaign budget.

How many rows can a poster comparison table hold?

This design is drawn for six rows and three columns, and five rows of short values sit comfortably in it at A0. Only three of the forty poster designs carry a table block at all, and their caps differ, so check the design before deciding how many models to compare.

Why does this figure have no legend?

The figure block on this design has its legend turned off, so the caption names the two bars instead. That is worth knowing, because on the designs that do show a legend it sits in the bottom margin under the axis, and long tick labels can collide with it. Short labels and short series names avoid that.

Does the layout have a references block?

This one does not. The forty poster designs differ in which blocks they carry, and here the three citations were written into the foot of the current practice section. That is the general fix: open the design, list its blocks, and move anything it lacks into a section it does have.

What does an A0 landscape poster export at?

The working canvas is 2,340 by 1,656 pixels and the export captures at twice that, so 4,680 by 3,312 pixels, roughly 100 dots per inch at full A0 size. Exports are RGB with no bleed setting, no crop marks and no CMYK conversion, so a print shop needing those will add them.

Build your own in about a minute

The button below opens the generator with this use case already described. Change the wording to match your own, generate, then edit anything you like.

Make one like this

Sources

Written and checked by the OneCraft team. Last checked 20 September 2026.