Booking Prediction
“Will this browsing session result in a booking?”
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
Will this browsing session result in a booking?
Online travel platforms convert only 2-4% of sessions into bookings. The 96-98% that don't convert represent a massive opportunity: even moving conversion from 3% to 3.5% is a 17% revenue increase. Traditional models use session-level features but miss the user journey graph: how search patterns evolve across sessions, how price sensitivity varies by trip context, and how property-user affinity signals predict intent. For an OTA with $5B in gross bookings, a 0.5% conversion lift generates $83M in incremental revenue.
Quick answer
Booking prediction AI identifies which browsing sessions have high conversion intent by analyzing the full user journey graph: search refinement patterns, engagement depth with property pages, user-property affinity, and loyalty history. Traditional conversion models use session-level features and miss the cross-session journey signals that predict intent. Graph-based models detect high-intent sessions in real time, enabling targeted interventions (urgency messaging, personalized offers) that lift conversion 0.5+ percentage points, generating $83M for an OTA with $5B in gross bookings.
Approaches compared
4 ways to solve this problem
1. Session-Level Heuristics
Flag sessions based on simple rules: viewed 3+ properties, spent 5+ minutes, reached the checkout page. Basic intent scoring available in most analytics platforms.
Best for
Quick-win conversion optimization when you have no ML infrastructure and need immediate signals.
Watch out for
Captures only the most obvious high-intent signals. Many high-intent users browse efficiently (few page views but high engagement per page), while low-intent users browse extensively (many views, low engagement each). Rules cannot distinguish these patterns.
2. Logistic Regression on Session Features
Build a conversion model using session features: page views, time on site, search filters used, device type. The standard approach for conversion optimization.
Best for
Platforms with clean session tracking and well-defined conversion funnels.
Watch out for
Session-level features miss the user journey context. A returning Gold loyalty member browsing Miami hotels for the 3rd time in a week has very different intent than a first-time visitor doing the same search. Also cannot capture user-property affinity: this user always books 4-star beach resorts at $250-350.
3. Deep Learning on Click Sequences (LSTM/Transformer)
Model the click sequence within a session as a time series and predict conversion from the evolving interaction pattern. Captures temporal dynamics better than flat features.
Best for
Platforms with rich clickstream data where the sequence of actions (search, filter, view, compare, view again) carries intent information.
Watch out for
Processes each session independently without user history context. Cannot represent that this user has been researching this trip across 5 sessions over 2 weeks, narrowing from 'beach vacation' to 'Miami, March 14-17, 4-star, ocean view.' The cross-session journey is the strongest intent signal, and sequence models within a single session miss it.
4. Graph Neural Networks (Kumo's Approach)
Connect users, searches, property views, bookings, and property attributes into a travel graph. GNNs learn booking intent from the full user journey, including cross-session patterns, user-property affinity, and loyalty context.
Best for
OTAs and hotel booking platforms with returning users, loyalty programs, and rich property catalog data where user-property affinity drives conversion.
Watch out for
Requires user identity across sessions (login or cookie matching). Anonymous first-time visitors with no history benefit less from the graph approach. Best value for platforms with 30%+ returning user traffic.
Key metric: Graph-based booking prediction identifies 2-3x more high-intent sessions than session-level models. A 0.5 percentage point conversion lift generates $83M for an OTA with $5B in gross bookings, a 17% relative revenue increase.
Why relational data changes the answer
Booking intent is built across a journey, not within a single session. User USR001 (Gold loyalty, 12 past bookings, $340 average) searching Miami hotels and spending 180 seconds on Ocean Breeze Resort viewing 8 photos is a very different signal than User USR002 (no loyalty, 0 bookings) spending 22 seconds on one property. The intent signal comes from the relational context: this user's booking history, their loyalty status, how this search compares to their typical booking pattern, and how deeply they engage with this specific property relative to their usual behavior.
Session-level models see two browsing sessions. Graph-based models see two nodes in a user-property-booking network, each with rich relational context that predicts intent with 80%+ accuracy for high-confidence segments. SAP's SALT benchmark shows graph models at 91% accuracy vs 63% for gradient-boosted trees on relational tasks. RelBench shows 76.71 vs 62.44 for GNNs. In booking prediction, this translates to identifying 2-3x more high-intent sessions, enabling targeted interventions (best-match recommendations, urgency messaging, loyalty-exclusive rates) that convert sessions that would otherwise bounce. A 0.5 percentage point lift on a 3% baseline is a 17% relative improvement in revenue.
Predicting booking intent from a single session is like a car salesperson judging a customer's intent from one showroom visit. They would miss that this customer has visited 4 dealerships this week, test-drove the same model at two competitors, and already has financing pre-approved. The purchase intent is built across the full journey, not visible in any single interaction. Graph-based booking prediction sees the full journey.
How KumoRFM solves this
Graph-powered intelligence for travel and hospitality
Kumo connects users, searches, property views, bookings, and property attributes into a travel graph. The GNN learns booking intent from the full user journey: how search refinement patterns signal high intent, how price sensitivity interacts with property attributes, and which user-property pairings have the highest conversion probability. PQL predicts booking probability per session, enabling real-time personalization and targeted incentives for high-intent sessions.
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
USERS
| user_id | loyalty_tier | past_bookings | avg_booking_value |
|---|---|---|---|
| USR001 | Gold | 12 | $340 |
| USR002 | None | 0 | N/A |
| USR003 | Silver | 5 | $220 |
SEARCHES
| search_id | user_id | destination | dates | guests | timestamp |
|---|---|---|---|---|---|
| SRC401 | USR001 | Miami | Mar 14-17 | 2 | 2025-03-01 10:00 |
| SRC402 | USR002 | NYC | Apr 5-7 | 1 | 2025-03-01 11:30 |
| SRC403 | USR003 | Miami | Mar 14-16 | 2 | 2025-03-01 14:00 |
VIEWS
| view_id | search_id | property_id | time_on_page_s | photos_viewed |
|---|---|---|---|---|
| VW601 | SRC401 | HTL001 | 180 | 8 |
| VW602 | SRC401 | HTL002 | 45 | 2 |
| VW603 | SRC402 | HTL003 | 22 | 1 |
BOOKINGS
| booking_id | user_id | property_id | total | timestamp |
|---|---|---|---|---|
| BK6001 | USR001 | HTL001 | $1,020 | 2025-03-01 10:25 |
PROPERTIES
| property_id | name | star_rating | avg_rate | review_score |
|---|---|---|---|---|
| HTL001 | Ocean Breeze Resort | 4-star | $295 | 4.6 |
| HTL002 | City Center Hotel | 3-star | $185 | 4.2 |
| HTL003 | Manhattan Suites | 4-star | $380 | 4.4 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT BOOL(BOOKINGS.booking_id, 0, 1, hours) FOR EACH SEARCHES.search_id
Prediction output
Every entity gets a score, updated continuously
| SEARCH_ID | USER_ID | DESTINATION | BOOKING_PROB | RECOMMENDED_ACTION |
|---|---|---|---|---|
| SRC401 | USR001 | Miami | 0.82 | Show best match |
| SRC402 | USR002 | NYC | 0.09 | Offer discount |
| SRC403 | USR003 | Miami | 0.44 | Show urgency |
Understand why
Every prediction includes feature attributions — no black boxes
Search SRC401 -- User USR001 searching Miami hotels
Predicted: 82% booking probability
Top contributing features
Detailed property review (180s + 8 photos)
High engagement
30% attribution
Loyalty tier and booking history
Gold, 12 past bookings
24% attribution
Date proximity (13 days out = committed)
Mar 14-17
19% attribution
Price alignment with avg booking value
$295 vs $340 avg
16% attribution
Search refinement pattern (narrowing)
2 destination searches
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 booking prediction
What is a good conversion rate for travel booking platforms?
Industry averages are 2-4% for OTAs, 3-6% for hotel direct booking sites, and 1-2% for metasearch. These rates have been remarkably stable for a decade despite massive investment in UX optimization, because the bottleneck is not the funnel but intent detection. Most platforms treat all sessions equally rather than concentrating conversion efforts on the 15-20% of sessions with genuine booking intent. Graph-based models identify these high-intent sessions and enable differentiated treatment.
How does booking prediction improve revenue without offering discounts?
The primary levers are: showing the best-matching property first (reducing search friction for high-intent users), displaying urgency signals ('only 2 rooms left at this rate') for users near the booking threshold, and reducing friction (pre-filling forms, offering one-click booking for loyalty members). These interventions do not reduce price. They reduce the effort and uncertainty that prevent high-intent users from completing their booking. Discounts are a last resort for moderate-intent users where a small price incentive tips the balance.
Can booking prediction work for first-time visitors with no history?
Partially. For anonymous first-time visitors, the model relies on within-session signals: search specificity (exact dates vs. flexible), engagement depth (time on page, photos viewed), and property-level conversion patterns (this hotel converts 12% of viewers vs. 3% for that hotel). Accuracy for first-time visitors is lower (55-65% vs 80%+ for returning users), but still valuable for the 70%+ of sessions that are anonymous. The gap narrows if you can match the user to a known device or email.
How does booking prediction integrate with real-time personalization?
The model outputs a booking probability score per session that updates in real time as the user interacts. This score drives personalization rules: sessions above 70% probability see best-match recommendations and streamlined checkout. Sessions at 30-50% see social proof ('42 people booked this hotel today') and urgency signals. Sessions below 20% see broader discovery experiences to build intent. The score is consumed by the personalization engine via API with sub-100ms latency.
What is the ROI of improving travel conversion rates?
For an OTA with $5B in gross bookings at 3% conversion, each 0.1 percentage point improvement generates $16.7M in incremental revenue. Graph-based models typically deliver 0.3-0.7 percentage point improvement, or $50-117M. Implementation costs are $1-3M including data integration and model development. The ROI is 20-100x, making conversion optimization one of the highest-return AI investments in travel. The revenue scales linearly with platform size.
Bottom line: An OTA with $5B in gross bookings generates $83M in incremental revenue by improving conversion 0.5 percentage points. Kumo's travel graph detects high-intent sessions from engagement depth, search refinement patterns, and user-property affinity signals.
Related use cases
Explore more travel & hospitality 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.
