samarv/oec-experimentation-framework

The Overall Evaluation Criterion (OEC) is a single composite metric used to determine the success of an experiment by balancing short-term gains against long-term health. It prevents teams from "gaming" the system by optimizing for narrow metrics that inadvertently hurt the user experience.

The OEC Construction Workflow

1. Identify Primary Success Metrics

Identify the immediate behavior you want to encourage (e.g., clicks, revenue, bookings). This is your "top-line" metric.

2. Identify Countervailing (Guardrail) Metrics

Determine what "bad" behaviors the top-line metric might encourage. These metrics act as a check on the primary goal.

• Revenue goals? Check for churn or user sentiment.
• Click-through goals? Check for bounce rate or "successful sessions" (e.g., clicks that stay on the page for >30 seconds).
• Email volume? Check for unsubscribe rates.

3. Model Lifetime Value (LTV)

Assign a dollar value to negative long-term actions to create a balanced formula.

• Example Calculation: OEC = (Short-term Revenue) - (Unsubscribe Rate * Estimated Value of a Subscriber).
• If an email campaign generates $1,000 but causes 500 unsubscribes, and each subscriber is worth $3/year, the campaign is net-negative (-$500).

4. Set Constraint Optimizations

If a composite formula is too complex, use the "Fixed Budget" approach.

• Allow the team to optimize a metric (e.g., Revenue) only if a guardrail metric (e.g., Page Load Latency or Ad Real Estate) remains within a fixed "budget."

Validity Testing with Twyman’s Law

"Any figure that looks interesting or different is usually wrong." Before celebrating a "home run" result:

• Check for Sample Ratio Mismatch (SRM): If you designed a 50/50 split but get 50.2/49.8 with 1M users, the experiment is likely invalid due to bot interference or data pipeline issues.
• Hold the Celebration: If a metric moves by 10% when you expected 1%, assume it is a bug (e.g., double-logging revenue) until proven otherwise.
• Replicate: Rerun the experiment. If the P-value stays significant (e.g., <0.01), trust increases.

The Experimentation Portfolio

Manage your product roadmap like a stock portfolio:

• Incremental (70-80%): Small, "inch-by-inch" improvements with high probability of success.
• High Risk/High Reward (10-20%): Radical redesigns or new features. Expect an 80-90% failure rate here, but aim for "home runs" that move the needle by 5-12%.

Examples

Example 1: Search Engine Monetization

• Context: A team wants to increase revenue by adding more ads to the search results page.
• Input: Primary metric = Ad Revenue. Countervailing metric = Session Success Rate (did the user find what they wanted?).
• Application: Create an OEC where revenue gains are only "wins" if the Session Success Rate does not drop by more than 0.1%.
• Output: The team realizes that while 4 ads make more money today, they increase churn. They settle on 3 ads as the optimal OEC balance.

Example 2: Retention Marketing

• Context: An email team is measured by "Revenue from Email Clicks."
• Input: Team increases email frequency to 5x per week.
• Application: Apply the "Cost of Spam" model. Subtract $5 from the "success" total for every unsubscribe.
• Output: The data shows that 50% of the high-frequency campaigns are actually destroying long-term value. The team reduces frequency to 2x per week, increasing total LTV.

Common Pitfalls

• Shipping on "Flat" Results: Never ship a feature if the results are not statistically significant. This introduces "code debt" and maintenance costs for zero user benefit.
• Ignoring the Sunk Cost Fallacy: Avoid shipping a 6-month project just because the team spent time on it. If the A/B test is negative, the project failed.
• Peeking at P-values: Do not stop an experiment early just because the P-value hit 0.05. This leads to "Type 1" errors (false positives). Wait for the full designated duration.
• Trusting One-off Successes: If a result is "too good to be true," it is likely a tracking bug or a bot. Apply Twyman's Law immediately.