Creative Performance Prediction
“Which creative will perform best for this segment?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
Which creative will perform best for this segment?
Creative testing is slow and expensive. Brands test 5-10 variants per campaign, wait 1-2 weeks for statistical significance, and still make suboptimal decisions because results don't generalize across segments. For an advertiser spending $30M on creative production and testing, predicting winners before launch saves $4-6M in wasted test spend and accelerates time-to-performance by 3x.
Quick answer
Graph neural networks predict which ad creatives will perform best for each audience segment by learning patterns across the full creative library, not just the current campaign. The model connects creative attributes (format, duration, CTA) to segment response histories, reducing creative testing cycles from weeks to hours and saving $4-6M in wasted test spend for a $30M advertiser.
Approaches compared
4 ways to solve this problem
1. A/B testing with statistical significance
Run 2-5 creative variants per segment, wait for enough impressions to reach 95% confidence, then pick the winner. The standard approach.
Best for
High-confidence decisions on individual campaigns. Straightforward to implement and explain.
Watch out for
Slow (1-2 weeks per test) and expensive (50-70% of test budget goes to losing variants). Results don't generalize across segments or campaigns, so you restart from scratch each time.
2. Multi-armed bandits (Thompson sampling, UCB)
Allocate traffic dynamically toward better-performing variants as results come in, reducing waste compared to fixed A/B splits.
Best for
Faster convergence than fixed A/B tests. Good for high-traffic campaigns where you can learn quickly.
Watch out for
Still learns from scratch per campaign. Cannot transfer knowledge from previous campaigns or predict performance for new creatives that haven't been tested yet.
3. Creative scoring models (computer vision + NLP)
Extract features from creative assets (colors, objects, text sentiment, video duration) and train a model to predict performance from these features.
Best for
Pre-launch screening of creative concepts. Can flag obviously poor creatives before any spend.
Watch out for
Treats each creative in isolation. Cannot model the interaction between creative attributes and audience segments. A video that works for Gen Z may fail for Boomers, and single-creative models miss this.
4. KumoRFM (relational graph ML)
Connect creatives, impressions, clicks, segments, and campaigns into a graph. The GNN learns which creative attributes resonate with which segments based on the entire historical creative library.
Best for
Predicts winning creatives before full deployment by transferring knowledge across campaigns, segments, and creative formats. Reduces test cycles from weeks to hours.
Watch out for
Requires historical performance data across a library of creatives and segments. Adds most value when you have 50+ historical creatives with segment-level performance data.
Key metric: Advertisers reduce creative testing cycles from 1-2 weeks to hours and save $4-6M on $30M creative budgets by predicting winners before full deployment.
Why relational data changes the answer
Creative performance is not a property of the creative alone. A 15-second video with a 'Shop Now' CTA performs differently for young professionals than for parents, differently on mobile than on TV, and differently in week one than in week four. Flat models that predict performance from creative features alone miss these interactions. They tell you 'video outperforms static' as a blanket rule, when the reality is 'video outperforms static for segment A on mobile in the first two weeks, but static outperforms video for segment B on desktop after the first week.'
Relational models connect creatives to segments to campaigns to impressions in a single graph. They learn these granular interaction patterns from your entire creative history, not just the current test. When you launch a new creative, the model already knows how similar creative attributes (format, duration, CTA type) performed for this segment on this publisher, producing a predicted performance score before a single impression is served. On the RelBench benchmark, relational approaches score 76.71 vs 62.44 for single-table baselines, and the gap is especially pronounced for interaction-heavy prediction tasks like creative-segment matching.
A/B testing creatives is like a restaurant that makes every diner taste-test five dishes before ordering. It works, but it wastes food and time. Graph-based creative prediction is like a chef who remembers that this diner likes spicy food, ordered the seafood last time, and that guests with similar taste profiles loved the new curry. The chef predicts the winning dish before the diner sits down.
How KumoRFM solves this
Graph-powered intelligence for advertising
Kumo connects creatives to impressions, clicks, segments, and campaigns in a graph that encodes creative attributes (format, length, CTA, color palette) alongside audience response patterns. The model learns which creative attributes resonate with which segments based on historical performance across the entire creative library, not just the current campaign.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
CREATIVES
| creative_id | format | duration | cta_type | campaign_id |
|---|---|---|---|---|
| CR001 | Video-15s | 15s | Shop Now | CMP01 |
| CR002 | Static | N/A | Learn More | CMP01 |
| CR003 | Video-30s | 30s | Sign Up | CMP02 |
IMPRESSIONS
| impression_id | creative_id | segment_id | timestamp |
|---|---|---|---|
| IMP701 | CR001 | SEG-A | 2025-02-20 |
| IMP702 | CR002 | SEG-A | 2025-02-20 |
| IMP703 | CR001 | SEG-B | 2025-02-21 |
CLICKS
| click_id | impression_id | timestamp |
|---|---|---|
| CLK501 | IMP701 | 2025-02-20 |
| CLK502 | IMP703 | 2025-02-21 |
SEGMENTS
| segment_id | name | size | avg_order_value |
|---|---|---|---|
| SEG-A | Young-professionals | 2.4M | $85 |
| SEG-B | Parents-35-50 | 3.1M | $120 |
CAMPAIGNS
| campaign_id | objective | budget | start_date |
|---|---|---|---|
| CMP01 | Conversions | $200K | 2025-02-15 |
| CMP02 | Sign-ups | $150K | 2025-02-18 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT COUNT(CLICKS.click_id, 0, 7, days) FOR EACH IMPRESSIONS.impression_id RANK TOP 1
Prediction output
Every entity gets a score, updated continuously
| SEGMENT_ID | CREATIVE_ID | PREDICTED_CTR | RANK |
|---|---|---|---|
| SEG-A | CR001 | 3.8% | 1 |
| SEG-A | CR002 | 1.9% | 2 |
| SEG-B | CR001 | 4.5% | 1 |
| SEG-B | CR003 | 2.1% | 2 |
Understand why
Every prediction includes feature attributions — no black boxes
Creative CR001 x Segment SEG-B (Parents-35-50)
Predicted: 4.5% predicted CTR (Rank #1)
Top contributing features
Video format preference for segment
2.3x vs static
32% attribution
15s duration engagement rate
87% completion
24% attribution
'Shop Now' CTA conversion history
3.1% avg
18% attribution
Segment purchase recency
12 days avg
15% attribution
Creative freshness (days since launch)
8 days
11% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about creative performance prediction
How do you predict ad creative performance before launch?
Connect your historical creative library to segment-level performance data in a graph model. The model learns which creative attributes (format, duration, CTA, visual style) perform best for which audience segments, enabling predicted CTR scores for new creatives before any impressions are served. This transfers learning across campaigns rather than starting from scratch each time.
What is the best way to optimize ad creatives?
Move beyond per-campaign A/B testing to cross-campaign learning. Graph models that connect creatives, segments, and performance data across your entire creative history can predict winners in hours instead of weeks. The key insight is that creative performance is a creative-segment interaction, not a creative-level property.
How do you reduce creative testing costs?
Pre-score creatives using a model trained on your historical creative library. This lets you eliminate obviously poor variants before spending any media budget and focus test spend on the top 2-3 candidates rather than 5-10. Advertisers spending $30M on creative production typically save $4-6M by reducing wasted test budget.
How does creative fatigue affect performance over time?
Creative fatigue is a time-dependent, segment-specific signal. A creative may fatigue quickly for a heavily-exposed segment while still performing well for a new audience. Graph models track fatigue per creative-segment-publisher combination, predicting when to rotate creatives before performance declines rather than reacting after CTR drops.
What data do you need for creative performance prediction?
Historical impression and click data with creative IDs, segment IDs, and timestamps. Creative metadata (format, duration, CTA type) and campaign context (objective, budget, publisher). The more creatives and segments in your history, the better the model predicts for new combinations.
Bottom line: An advertiser spending $30M on creative production saves $4-6M by predicting winning creatives before full deployment. Kumo's graph connects creative attributes to segment response patterns across the entire historical library, reducing test cycles from weeks to hours.
Related use cases
Explore more ad tech use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
