Case context: You are building two predictive systems: one that sorts customer support emails into "urgent" and "non-urgent," and another that forecasts next-day prices for a highly volatile cryptocurrency using long historical sequences of market data. You want to estimate the best achievable error rate for each system to support bias and variance analysis.

Question: For the two applications described, explain whether the best achievable error rate is practical to estimate and how your method should differ across the two tasks.

Sample answer: The email-sorting task is relatively human-friendly because people can usually judge urgency with reasonable accuracy. For that reason, human performance can serve as a useful proxy for the best achievable error rate. By contrast, forecasting a volatile cryptocurrency price is much harder for people to do well, so human performance is not a reliable proxy. For that task, estimating the best achievable error rate is much more difficult and may be impossible to do accurately.

Key points:

• Email sorting is human-friendly, so human performance can estimate the best achievable error.
• Cryptocurrency price forecasting is not human-friendly, so the best achievable error is hard to estimate.
• The type of task determines how to approximate the best achievable error rate.

Rubric: The student must correctly identify that the email sorter is a human-friendly task where human performance can estimate the best achievable error rate, while the cryptocurrency forecaster is a task humans struggle with, making the best achievable error rate very difficult to estimate.