JPMorgan Data Science Analyst Interview Guide: Technical Depth Without Overcomplicating

What the Interview Usually Tests| 这类面试一般会测哪几件事

• Statistical reasoning: experimental design, sampling bias, probability, and model evaluation.

  统计判断：实验设计、抽样偏差、概率和模型评估。

• Modelling judgement: when to choose a simpler model, how to handle imbalance, and what to monitor after launch.

  建模判断：什么时候该用简单模型、类别不平衡怎么处理、上线后看哪些监控项。

• Python or SQL fluency: data cleaning, aggregation, feature preparation, and result checking.

  Python 或 SQL 基础：数据清洗、聚合、特征准备和结果校验。

• Business communication: what the result means for fraud, risk, operations, or customer decisions.

  业务沟通：结果对风控、运营、客户决策到底意味着什么。

Statistics Questions Are Usually About Judgement| 统计题本质上常常是在考判断

• If you mention an A/B test, also mention sample ratio mismatch, leakage, and how you would define the success metric before running it.

  如果你提到 A/B test，就顺手把 sample ratio mismatch、数据泄漏，以及实验前先定义成功指标这些点带出来。

• If you mention precision and recall, explain which mistake is more expensive in the business context.

  如果你讲 precision 和 recall，就继续说在当前业务里哪种错误更贵。

• If you mention a model uplift, say what the operational action would be once the score is available.

  如果你提到模型带来的 uplift，也要说分数出来以后业务动作是什么。

How to Talk About Modelling Like a Practitioner| 怎么把建模题讲得像真正做过的人

• Start from the decision, not the algorithm. What business action is the model supporting?

  从决策出发，不是从算法出发。这个模型到底是为了支持什么业务动作？

• Explain your baseline. Interviewers trust candidates who compare against a simple alternative.

  先讲 baseline。愿意和简单方案比较的候选人，通常更让人信任。

• Name one likely failure mode such as drift, proxy leakage, or unstable labels.

  主动说一个常见失败点，比如 drift、proxy leakage 或标签不稳定。

• Describe how you would monitor post-launch quality instead of pretending the model is done once it ships.

  说明上线后怎么监控，而不是把模型讲成上线就结束的静态项目。

Python and SQL: Keep It Clean and Explain the Intention| Python 和 SQL 题：代码写清楚，意图讲清楚

import pandas as pd

summary = (
    df.groupby("variant")
      .agg(users=("user_id", "nunique"), conversions=("converted", "sum"))
      .assign(conversion_rate=lambda x: x["conversions"] / x["users"])
      .reset_index()
)

print(summary)

What to Do When You Do Not Know the Exact Answer| 遇到不会的技术题，怎么答才不崩

• State the part you are sure about before guessing the rest.

  先说你确定的那一部分，再谈不确定的部分。

• Offer a sensible fallback or approximation instead of freezing.

  给出一个靠谱的替代思路，不要直接卡死。

• Say what you would validate with more time or data.

  说明如果时间和数据更充足，你下一步会验证什么。

• Keep the answer business-linked. Technical uncertainty is easier to forgive than losing the plot entirely.

  把答案和业务目标连回去。技术点一时不全没关系，彻底失去问题主线才更致命。

A Focused One-Week Plan| 一个更聚焦的一周准备方案

• Days 1-2: Review probability, experiments, sampling, and model metrics with small examples you can explain aloud.

  第 1 到 2 天：复习概率、实验、抽样和模型指标，重点不是看懂，而是能开口解释。

• Days 3-4: Practise five Python or SQL tasks end to end and narrate your choices as you code.

  第 3 到 4 天：练 5 道 Python 或 SQL 题，从头到尾做完，并边写边讲你的判断。

• Day 5: Rehearse model design questions using one fraud, one customer, and one operations scenario.

  第 5 天：用一个风控场景、一个客户场景、一个运营场景，练模型设计题。

• Days 6-7: Do mock interviews that force you to switch between statistics, code, and business explanation.

  第 6 到 7 天：找人做 mock，要求你在统计、代码和业务解释之间来回切换。