Describe EDA for a 3-year daily user sign-up dataset.

WHAT THIS TESTS: This question tests your ability to perform structured, first-principles Exploratory Data Analysis (EDA) on time series data. The interviewer isn't looking for complex forecasting models. They want to see if you can methodically decompose a time series into its fundamental components: trend, seasonality, and noise. It's a test of foundational data literacy and a systematic approach to problem-solving before jumping to conclusions.

A GOOD ANSWER COVERS: A strong answer discusses the four key components to investigate, in order. First, the TREND: the long-term increase, decrease, or stagnation of sign-ups over the three-year period. Is the business growing? Second, SEASONALITY: fixed, periodic fluctuations. This includes weekly patterns (e.g., fewer sign-ups on weekends) and yearly patterns (e.g., a dip in December, a spike in January). Third, CYCLICAL PATTERNS: longer-term, non-fixed fluctuations, which might be tied to broader economic conditions. For a 3-year dataset, this is less likely but shows advanced understanding. Fourth, IRREGULARITY / ANOMALIES: unpredictable spikes or dips that don't fit the other patterns. These often correspond to specific events like bot attacks, viral marketing, or site outages and require investigation.

COMMON WRONG ANSWERS: The most common mistake is jumping directly to forecasting. Mentioning ARIMA, Prophet, or LSTM models before describing the data's characteristics is a major red flag. It shows you're focused on tools, not on understanding the data first. Another weak answer is being too simplistic, for example, only stating "I'd check if sign-ups are going up." This misses the rich, actionable detail in seasonality and anomalies. Finally, confusing seasonality (fixed, e.g., every Monday) with cycles (variable, e.g., business cycles) is a sign of a less experienced candidate.

LIKELY FOLLOW-UPS: Expect questions like: "You see a huge, one-day spike. How would you investigate its cause?" (Answer: check marketing campaign launches, press mentions, referral logs, engineering incident reports). Or, "How would you quantify the weekly seasonality?" (Answer: group by day of the week and compare averages, or use an ACF plot to show correlation at a 7-day lag). Another is, "How would you report on growth, accounting for the Q4 holiday dip?" (Answer: use Year-over-Year growth for the same month/quarter).

ONE CONCRETE EXAMPLE: For a 3-year daily sign-up dataset, I would first plot the raw data to visually inspect the long-term trend. I'd expect to see growth. Then, I'd look for seasonality by plotting the average sign-ups by day of the week, likely finding a 20-30% drop on Saturday/Sunday. For yearly seasonality, I'd plot an overlay of each year's data, expecting to see a dip around late December and a spike in early January. A single point on May 1st, 2022 that is 5x the rolling average is an anomaly. I would query our marketing database to see if a promotion started that day. The remaining random variation is the noise.