Time-series interview questions on leakage, seasonality and backtests

Concepts

Q1. What makes a time series different from ordinary tabular data?

Rows are ordered and usually dependent: today's value is related to yesterday's. That breaks the assumption behind random cross-validation and means the error of one forecast depends on how far ahead it looks.

Listening for: dependence and ordering, not only "it has a date column".

Q2. Trend, seasonality and cycles — the difference?

Trend is a long-run direction. Seasonality repeats at a fixed period — day of week, month, festival calendar. Cycles rise and fall without a fixed period, such as business cycles. A daily sales series in India might show weekly seasonality, an annual festival peak and a trend from store expansion.

Q3. What is stationarity and why does it matter?

A stationary series has a stable mean, variance and autocorrelation over time. Classical models such as ARIMA assume it after differencing. Machine learning models do not formally require it, but a strongly trending target makes tree models fail at new levels they never saw, which is why forecasting the change or ratio is often safer.

Q4. What does an autocorrelation plot at lag 7 tell you for daily data?

A high autocorrelation at lag 7 (and 14, 21) indicates weekly seasonality: each day resembles the same weekday a week earlier. It suggests a seasonal-naive baseline and day-of-week features.

Leakage

Q5. Name three ways leakage enters a forecasting pipeline.

1. 1Random train-test splits, which train on the future.
2. 2Rolling features computed with a centred window, or including the current day's value when forecasting that day.
3. 3Joining an external table by date where the value was only published later — for example, a monthly figure released three weeks after month end.

Listening for: the third one. It is the hardest to spot and the one real pipelines get wrong.

Q6. How do you build a 7-day rolling mean feature safely?

Shift before rolling, so the feature for day t uses days t−7 to t−1:

import pandas as pd

s = pd.Series([10, 12, 11, 13, 30, 12, 14, 15],
              index=pd.date_range("2026-03-01", periods=8, freq="D"))
safe = s.shift(1).rolling(3).mean()      # window of 3 for a short example
leaky = s.rolling(3).mean()              # includes the value being predicted
print(pd.DataFrame({"y": s, "safe_mean3": safe, "leaky_mean3": leaky}).round(2))

On 5 March the leaky feature (18.0) already contains the spike of 30 it is supposed to predict; the safe feature (12.0) does not. If you forecast h days ahead, shift by h, not 1.

Baselines and metrics

Q7. What baseline would you beat before trying a model?

A naive forecast (tomorrow equals today) and a seasonal-naive forecast (tomorrow equals the same weekday last week). If a model cannot beat seasonal naive on a proper backtest, it is not ready.

Q8. Why is MAPE risky?

It divides by the actual value, so it explodes for small actuals and is undefined at zero, and it penalises over-forecasts and under-forecasts asymmetrically. For intermittent demand, use MAE, a scaled error such as MASE, or weighted MAPE computed as total absolute error divided by total actuals.

Backtesting

Q9. Describe a rolling-origin backtest.

Pick several forecast origins. At each, train only on data up to the origin, forecast the next h periods, and record the error. Move the origin forward and repeat. The average error across origins estimates real performance, and the spread shows how stable it is.

The block below backtests naive and seasonal-naive forecasts on four weeks of hypothetical orders with weekly seasonality:

import pandas as pd

weekly_pattern = [120, 100, 95, 98, 110, 160, 175]   # Mon..Sun, hypothetical
orders = pd.Series(
    [v + 2 * week for week in range(4) for v in weekly_pattern],
    index=pd.date_range("2026-02-02", periods=28, freq="D"),   # starts on a Monday
)

horizon, results = 7, []
for origin in [14, 21]:                        # train on the first 14, then 21 days
    train, test = orders.iloc[:origin], orders.iloc[origin:origin + horizon]
    naive = pd.Series(train.iloc[-1], index=test.index)
    seasonal = pd.Series(train.iloc[-7:].values, index=test.index)
    results.append({"origin": str(orders.index[origin].date()),
                    "naive_MAE": (test - naive).abs().mean(),
                    "seasonal_MAE": (test - seasonal).abs().mean()})
print(pd.DataFrame(results).round(1).to_string(index=False))

Naive error is 51 orders at both origins, because the last training day is a Sunday, the busiest day, and its value is carried across the whole following week; seasonal naive misses only by the trend of 2 orders per week. That gap is what the interviewer wants you to explain.

Q10. How long should the test window be, and how many origins?

At least the forecast horizon, and enough origins to cover the seasonal cycle that matters — for retail with festival peaks, origins across more than one festival season if the history allows. One test window around a festival can make any model look good or bad by chance.

Scenario

Q11. "Our forecast was fine all year and badly wrong in the Diwali week. What do you do?"

Check whether the model knows about the festival: a fixed calendar-month feature misses a festival that moves between October and November, so add a holiday feature built from the actual dates. Compare the error in last year's festival week in the backtest; if it was also poor, the problem is structural. Consider separate handling for event periods — an uplift factor estimated from past festivals applied to the base forecast — and report event-week error separately so a good yearly average does not hide it.

How to practise

Take any daily series — orders from a practice dataset, or public weather records — and write the Q9 backtest with three forecast origins, then add one model and see whether it beats seasonal naive. Extending feature engineering with pandas to lag features is the natural next step, and error metrics are compared in the model evaluation bank. Forecasting with proper backtests is taught in the Data Science programme.