User Intent Prediction #9: Time-Series Motivation. Model Said No.
I tried to turn a snapshot into a movie. The critics hated it.
I had a magic number.
My hand-crafted motivation_score (Article #2) had 0.96 prediction accuracy.
I thought: “If the current score is this good, surely the history of the score is even better?”
- Is the user getting more motivated?
- Are they losing interest?
- The trajectory must matter!
I spent 3 days building an sequence model to model the “motivation journey.”
The Experiment
I fed the model a sequence of motivation scores over time.
- Screen 1: Motivation = 0.2 (Browsing)
- Screen 5: Motivation = 0.4 (Curious)
- Screen 10: Motivation = 0.8 (Engaged)
- Screen 15: Motivation = 0.9 (Ready to buy)
The sequence model would learn the “arc” of the user’s journey. It would see patterns like:
- “Slow rise → Buyer”
- “Sharp spike → Impulse buyer”
- “Plateau → Window shopper”
The Theory: The trajectory tells a story. The story predicts the sale. The Result: The model didn’t care about the story.
The Showdown
| Feature | Prediction Quality |
|---|---|
| Current Motivation Score | 0.960 |
| Motivation Trajectory (sequence model) | 0.961 |
I added a neural network. I added complexity. I added history. I gained 0.001 prediction accuracy.
That’s not a win. That’s a rounding error.
The “Markov Property” of Intent
Why didn’t the history matter?
Because intent is a state, not a story.
Example:
- User A: Bored for 10 minutes (0.2), then suddenly engaged (0.9) at the paywall.
- User B: Engaged for 10 minutes (0.9), then suddenly bored (0.2) at the paywall.
The sequence model saw two different “arcs.” But only the final state mattered.
If a user is highly motivated right now (at the paywall), it doesn’t matter if they were bored 5 minutes ago. If a user is bored right now, it doesn’t matter if they were excited 5 minutes ago.
The current moment is everything.
I was trying to model a “narrative arc” of the user. But users don’t have arcs. They have impulses. The impulse to buy exists in the present tense.
The Lesson
Don’t over-engineer the obvious.
I had the answer (0.96 prediction accuracy). But I couldn’t leave well enough alone. I had to add “AI” to it.
I tried to turn a snapshot into a movie. But the sale happens in a snapshot.
History is irrelevant when the credit card is in hand.
This leads to finding universal patterns across cohorts—if history doesn’t matter, maybe demographics don’t matter either?