Full Case Study
PRD: Automated Pattern Detection Engine
A full walkthrough of how I scoped, prioritized, and shipped Chart AI's most-used feature — from trading pain to shipped product.
PRD_AutoPattern_Detection_v2.md
PRD: Automated Pattern Detection Engine
Author: Areeb Ali
Status: Shipped
Version: 2.1
Product: Chart AI · iTechGemini
Overview
ProblemTraders miss chart patterns due to screen fatigue and multi-timeframe complexity
GoalAutomate recognition of 12 high-probability chart patterns across 3 timeframes
Success Metric≥70% accuracy vs manual ID; 20%+ DAU increase within 30 days of launch
PriorityP0 — Core feature for Q1 roadmap HIGH
StakeholdersProduct (me), Engineering, QA — sole domain expert driving requirements
Background & Motivation
This feature originated directly from my own trading experience. Scanning 20+ charts across multiple timeframes manually takes 2–3 hours — by the time a pattern is identified, the setup may have already played out. The gap: no mobile-native app offered reliable automated detection with enough accuracy to trust. I validated this gap personally across 8 competitor apps before writing a single line of spec.
User Pain Points (from trading experience)
- Scanning 20+ charts manually takes 2–3 hours — patterns missed by the time user acts
- Most pattern tools use lagging indicators — signals arrive after the move, not before
- No mobile-native solution existed — all existing tools were desktop-only adaptations
- False positives erode trust faster than missed signals — accuracy threshold is non-negotiable
- Alert fatigue from low-quality signals causes users to disable notifications entirely
- No confidence scoring — users couldn't differentiate strong setups from weak ones
Feature Scope
In ScopeHead & Shoulders, Double Top/Bottom, Bull/Bear Flags, Ascending/Descending Triangles, Wedges, Cup & Handle
Out of ScopeElliott Wave (too subjective), Harmonic patterns (v3 candidate — accuracy insufficient at v2)
Timeframes1H, 4H, 1D — highest signal-to-noise ratio per 5+ yrs trading data
AssetsTop 50 by market cap + user watchlist assets
Excluded TFs15m, 5m, 1m — retail traders lose money on low TF noise; excluded by design
Acceptance Criteria
- Pattern overlay renders within 800ms of symbol load MUST
- Confidence score displayed (Low / Medium / High) with color coding MUST
- Minimum 70% accuracy vs manual identification in QA testing MUST
- Pattern name, description, and typical price target visible on tap MUST
- Push notification on new pattern detection (opt-in, max 3/day per asset) SHOULD
- Historical hit rate visible per pattern type across 90-day lookback NICE
- Pattern filtering by confidence level NICE
Domain Rationale (the trading perspective non-traders miss)
- 1H/4H/1D chosen over lower TFs — retail traders are trend-followers, not scalpers. Sub-1H patterns have poor R:R for mobile users
- Confidence score prevents over-reliance — trading is probabilistic. A "High" confidence label sets context, not certainty
- Excluded harmonic patterns — Gartley/Bat/Crab patterns require precise Fibonacci ratios; auto-detection false positive rate was too high in testing
- Cup & Handle included despite complexity — one of the highest win-rate patterns in trending markets; worth the engineering effort
- Max 3 push notifications/day per asset — prevents alert fatigue. I tested higher frequencies personally: >3 causes users to mute all alerts within a week
Risk & Mitigation
- Risk: False positives damage user trust → Mitigation: Hard gate at 70% accuracy; QA validates vs my manual identification on 200 live charts
- Risk: Users over-trade based on signals → Mitigation: Confidence scoring + educational tooltip on every pattern ("this is context, not advice")
- Risk: Engineering complexity delays timeline → Mitigation: Phased rollout — 6 patterns at launch, 12 by v2.1
⚠ The Problem I Lived
I spent 2–3 hours daily scanning charts manually as part of my trading routine. The problem wasn't motivation — it was physics. No human can reliably identify patterns across 20+ assets and 3 timeframes simultaneously. The cognitive load creates blind spots, and missed patterns mean missed trades. I knew exactly what the tool needed to do because I was the user who needed it.
✓ My Product Approach
I wrote the spec as the domain expert, not just as a PM. Every requirement was pressure-tested against live trading sessions. The 70% accuracy threshold wasn't arbitrary — it was the minimum where I'd trust a signal myself. The 800ms load time requirement came from knowing traders won't wait: if the chart loads slowly, they switch apps mid-trade.
Product Decisions
The calls that shaped the feature
Why automated analysis over more chart types?
After analyzing user behavior, traders spent more time seeking entry signals than exploring chart layouts. Automated pattern recognition delivers repeated daily value — chart type variety is a one-time choice. Signal generation keeps users coming back. The retention data confirmed this: pattern detection drove the biggest single DAU lift since launch.
How did I validate before committing dev resources?
I stress-tested every proposed feature against my own trading sessions first. If I couldn't see myself using it while a trade is live, it didn't make the roadmap. For pattern detection specifically, I ran a manual "simulation" for two weeks — identifying patterns by hand, noting time cost — before writing the first spec line. This eliminated 3 low-value pattern types from scope before dev saw them.
What was the hardest product tradeoff?
Balancing simplicity for retail vs depth for experienced traders. Solved with progressive disclosure — clean defaults with advanced layers accessible but not mandatory. The confidence score system was the key unlock: new traders see "High / Medium / Low" without being overwhelmed, while experienced traders understand the probabilistic framework underneath.
Why cap push notifications at 3/day?
I tested this personally. More than 3 notifications per asset per day trains users to ignore them. The marginal value of the 4th alert is negative — it increases noise perception. This was a non-obvious call that came from personal experience with alert fatigue, not from any user research doc.
"The best crypto products are built by people who use them. I'm one of the few PMs who can validate a feature in the morning and ship a spec by afternoon — because I tested it in a live trade."
— Areeb Ali, on why domain expertise is a product moat
Outcomes
What shipped and what it produced
100K+
Total downloads (product)
12
Patterns shipped by v2.1
≥70%
Accuracy threshold hit
Lessons Learned
Traders need signal, not noise
Domain expertise is a product moat
Validate in live markets before spec
Progressive disclosure unlocks both cohorts
Alert frequency matters as much as quality
Accuracy gates protect long-term trust
Ship fast, compound gains