▸case-01 We are building a demand forecasting model using 3 years of daily store transaction data. A team member suggests using standard 10-fold random cross-validation to assess model performance across randomly shuffled folds. Explain why random K-fold is flawed for this scenario and specify the appropriate cross-validation technique to prevent data leakage. | pass→pass | 11,155 | 10,205 | -9% | 1 | 1 | 0% | 2,014 | 1,781 | -12% | 0 | 0 | — |
▸case-02 We are evaluating a daily inventory forecasting model for spare parts where many days have zero sales. The product manager proposes using Mean Absolute Percentage Error (MAPE) to report accuracy to stakeholders. Explain the mathematical failure mode of MAPE in this context and specify a scale-aware error metric suited for intermittent time series. | fail→pass | 20,721 | 12,085 | -42% | 1 | 1 | 0% | 2,732 | 2,295 | -16% | 0 | 0 | — |
▸case-03 We have monthly electricity consumption data spanning 10 years showing strong upward growth. Before fitting an autoregressive integrated moving average model, how should we statistically test whether the series is stationary, and what transformation should be applied if the test indicates a unit root? | pass→pass | 13,286 | 12,193 | -8% | 1 | 1 | 0% | 2,501 | 2,382 | -5% | 0 | 0 | — |
▸case-04 An analyst is analyzing monthly web traffic data that exhibits changing seasonal amplitude and non-linear trend over 5 years. They plan to use classical additive decomposition. Explain why classical decomposition is inadequate for changing seasonal patterns and specify an alternative robust decomposition algorithm. | pass→pass | 15,187 | 10,085 | -34% | 1 | 1 | 0% | 2,139 | 1,762 | -18% | 0 | 0 | — |
▸case-05 When auto-fitting an ARIMA model on a stationary univariate series, a developer suggests picking order parameters (p, d, q) by selecting the combination that maximizes R-squared on the training data. What is the standard statistical selection criterion used to balance goodness-of-fit and model complexity without overfitting? | pass→pass | 6,601 | 7,326 | +11% | 1 | 1 | 0% | 1,249 | 1,424 | +14% | 0 | 0 | — |
▸case-06 We observe a strong 12-month recurring spike in monthly retail sales data. To achieve stationarity for a seasonal ARIMA model, an engineer applies simple lag-1 differencing (y_t - y_{t-1}). Why is lag-1 differencing insufficient for strong annual seasonality, and what differencing operation should be performed? | pass→pass | 12,210 | 8,608 | -30% | 1 | 1 | 0% | 1,735 | 1,645 | -5% | 0 | 0 | — |
▸case-07 We are engineering lag features for a machine learning model predicting hourly traffic flow. An analyst suggests computing a 24-hour centered moving average as a feature for predicting traffic at time t. Explain why this feature violates forecasting constraints and state how to correctly construct moving average features. | pass→pass | 9,226 | 9,763 | +6% | 1 | 1 | 0% | 1,704 | 1,861 | +9% | 0 | 0 | — |
▸case-08 After fitting a SARIMAX model to quarterly revenue, we need to verify if the model residuals represent white noise. A junior analyst calculates the Pearson correlation between residuals and time. Explain why Pearson correlation is inappropriate for residual diagnostics and specify the statistical test designed to check residual autocorrelation across multiple lags. | pass→pass | 10,758 | 11,034 | +3% | 1 | 1 | 0% | 1,936 | 1,964 | +1% | 0 | 0 | — |
▸case-09 A telemetry dataset records server memory utilization, but timestamps are irregularly spaced due to event-driven logging (ranging from 2 seconds to 15 minutes apart). An analyst wants to fit a standard ARIMA model directly on these raw sequential rows. What preprocessing step must be performed before applying standard discrete-time time series models? | pass→pass | 6,282 | 4,752 | -24% | 1 | 1 | 0% | 1,090 | 836 | -23% | 0 | 0 | — |
▸case-10 We want to model the dynamic interactions between inflation rate, interest rate, and unemployment rate over time. A practitioner proposes fitting a Vector Autoregression (VAR) model directly on raw, un-differenced macroeconomic level series that exhibit unit roots. What essential check and transformation must be performed on each series prior to VAR estimation? | pass→pass | 8,879 | 7,032 | -21% | 1 | 1 | 0% | 1,666 | 1,413 | -15% | 0 | 0 | — |
▸case-11 Daily stock returns exhibit periods of low volatility followed by clusters of high volatility. A researcher attempts to model this using standard linear OLS regression with constant error variance assumptions. What family of time series models specifically addresses conditional heteroskedasticity and volatility clustering in financial returns? | pass→pass | 6,919 | 4,747 | -31% | 1 | 1 | 0% | 1,295 | 950 | -27% | 0 | 0 | — |
▸case-12 Before deploying a complex neural network for daily energy load forecasting, the team needs a standard zero-parameter baseline model to evaluate whether the complex model adds value. An engineer suggests using a random walk with zero trend as baseline. What standard baseline forecasting method uses the most recent observation (or same period last year) as the predicted value? | pass→fail | 4,944 | 4,936 | -0% | 1 | 1 | 0% | 950 | 944 | -1% | 0 | 0 | — |
▸case-13 An algorithmic trading strategy monitors two asset price series that are both non-stationary I(1) processes. A trader proposes calculating Pearson correlation on raw prices to detect trading signals. Explain why raw price correlation can lead to spurious regression and identify the statistical framework used to test for a stationary linear combination between two I(1) series. | pass→pass | 14,272 | 12,704 | -11% | 1 | 1 | 0% | 2,706 | 2,362 | -13% | 0 | 0 | — |
▸case-14 Astronomers observing variable star brightness have non-uniformly spaced night observation times. An engineer attempts to apply standard Fast Fourier Transform (FFT) directly to the un-grid data. Why does standard FFT fail here, and what periodogram algorithm is designed for unevenly sampled time series? | pass→pass | 12,097 | 11,611 | -4% | 1 | 1 | 0% | 2,080 | 2,049 | -1% | 0 | 0 | — |
▸case-15 We are forecasting monthly gas demand and want to include local temperature forecasts as an exogenous predictor in our time series model. An analyst suggests fitting standard linear OLS regression on sales vs temperature. Why does standard OLS fail on time series data with autocorrelated residuals, and what model extends ARIMA to handle exogenous regressors? | pass→pass | 13,430 | 13,830 | +3% | 1 | 1 | 0% | 2,347 | 2,517 | +7% | 0 | 0 | — |
▸case-16 We are analyzing quarterly hotel occupancy rates which show linear growth trend and seasonal fluctuations whose magnitude increases proportionally with the overall level of the series. Which variant of Holt-Winters exponential smoothing (additive vs multiplicative seasonality) should be applied? | pass→pass | 4,068 | 4,090 | +1% | 1 | 1 | 0% | 729 | 709 | -3% | 0 | 0 | — |
▸case-17 To forecast 12 months ahead, a team compares recursive multi-step forecasting (feeding model predictions back as input lags) against direct multi-step forecasting (training 12 separate models, one for each horizon). What is the primary drawback of the recursive strategy for long forecast horizons? | pass→pass | 4,446 | 4,817 | +8% | 1 | 1 | 0% | 693 | 700 | +1% | 0 | 0 | — |
▸case-18 A retailer wants to summarize forecasting accuracy across 10,000 SKUs ranging from $1 items to $5,000 appliances. The data scientist proposes Mean Absolute Error (MAE) aggregated across all SKUs. Explain why raw MAE is inappropriate for cross-item comparison and specify a scale-independent metric. | pass→pass | 13,611 | 11,129 | -18% | 1 | 1 | 0% | 2,643 | 2,088 | -21% | 0 | 0 | — |
▸case-19 To detect real-time anomalies in network bandwidth usage, an engineer uses static global thresholding (flagging values > 3 standard deviations from the overall historical mean). Why does global static thresholding produce false alarms during scheduled high-traffic hours, and what dynamic window technique should be used? | pass→pass | 10,167 | 10,263 | +1% | 1 | 1 | 0% | 1,806 | 1,826 | +1% | 0 | 0 | — |
▸case-20 We have a single customer database table containing demographic snapshot attributes (age, income, region, total spend over past year). The marketing team wants to segment customers into 5 groups for targeted campaigns using K-Means clustering. How should we configure time series stationarity tests and lag features for this customer segmentation task? | pass→pass | 11,349 | 8,647 | -24% | 1 | 1 | 0% | 2,058 | 1,640 | -20% | 0 | 0 | — |
▸case-21 In a medical study, 500 patients are tracked to record time-to-event data (months until disease relapse or study dropout). A biostatistician considers using ARIMA forecasting on the patient duration numbers. Explain why ARIMA is inappropriate for right-censored time-to-event survival data and specify the standard statistical model for estimating hazard ratios. | pass→pass | 10,953 | 10,221 | -7% | 1 | 1 | 0% | 2,075 | 1,847 | -11% | 0 | 0 | — |
▸case-22 A mining company collected topsoil lead concentration measurements at 200 geographic GPS coordinate locations across a county at a single point in time. An analyst suggests fitting a SARIMAX time series model to map contamination across the county. Why is time series modeling applicable here, and what spatial interpolation technique should be used? | pass→pass | 11,369 | 9,792 | -14% | 1 | 1 | 0% | 2,102 | 1,742 | -17% | 0 | 0 | — |