Forecasting
54 bites tagged Forecasting — interview questions with model answers, and 60-second explainers.
Capacity planning for a distributed cache
Track hit ratio, memory and eviction rate, throughput/latency, and connections; correlate with growth to forecast when to add capacity before the hit ratio or evictions… Whether you can pick the right cache metrics and forecast scaling.
Communicate a forecast interval to an executive
Give the point estimate but frame the range as scenarios, use a fan chart, tie the interval to planning decisions and risk. communicating uncertainty to leadership. presenting $10M as a guaranteed single number with no range.
Reframe time series for a tree model
Lag and rolling-window features, calendar and cyclical encodings, then split chronologically to avoid leakage. turning forecasting into supervised learning. random shuffling that lets future data leak into training.
Sensitivity analysis on a growth model
Perturb each input by a normalized amount, measure the change in the long-term output, and use elasticities or global methods to rank drivers. finding leverage in a coupled model.
Communicate forecast uncertainty with prediction intervals
A point estimate hides risk; produce a prediction interval via model error, simulation, or scenarios, and state assumptions. quantifying and communicating forecast uncertainty.
Stationarity in time series and why ARIMA needs it
Constant mean/variance/autocovariance; ARIMA's coefficients assume them; test with the ADF test and ACF plots; achieve it via differencing or log transforms. whether you know stationarity means stable statistical properties over time.
Which classical baseline model handles weekly seasonality and upward trend?
Tests matching model structure to data characteristics. Name Holt-Winters triple exponential smoothing; map its level, trend, and seasonal equations to weekly period. Red flag: jumping to SARIMA without explaining why ETS is the natural baseline.
How do you measure forecast accuracy and compare MAE to RMSE?
This tests out-of-sample validation and how MAE and RMSE weight errors. A strong answer demands a train-test split, defines both, and notes RMSE punishes outliers more while MAE is more robust. A red flag is citing in-sample fit instead of held-out error.
Describe a strategy for reconciling different forecasts into one robust prediction
Tests synthesis of heterogeneous models into a consensus forecast. Strong answers diagnose divergence drivers first, then weight by track record or uncertainty, and output a distribution. Red flag: blind averaging without understanding why models disagree.
How would you incorporate external events into a time-series forecast?
This tests exogenous regressor design. A strong answer proposes binary or continuous regressors for holidays and campaigns in Prophet or ARIMA, then validates lift via backtesting. A red flag is dropping outlier days or applying post-hoc manual adjustments.
Why is extrapolating 5% weekly growth naive for annual forecasting?
This tests knowledge of extrapolation limits. A strong answer cites market saturation, seasonality, and channel exhaustion as invalidating factors, and notes that compounding 5% over 52 weeks magnifies error.
What data and approach for a simple 30-day DAU forecast?
Tests forecasting from sessionized logs without overengineering. Cite timestamped events, a 30 min session rule, and a regression baseline with day-of-week, recent totals, scored with MAE. Red flag: deep learning before a baseline or ignoring privacy hashing.
How do you prevent future leakage in time-series preprocessing?
This tests temporal causality in feature engineering and validation. Use only past data for lags and rolling windows and enforce a rolling validation split without shuffling. Red flags are random k-fold CV and global standardization leaking future information.
Describe two methods for generating prediction intervals or probabilistic forecasts
Tests uncertainty quantification for risk-adjusted decisions. Strong answers: (1) parametric intervals via forecast error variance and normal multipliers, (2) bootstrap residual resampling for empirical percentiles.
Random split vs walk-forward validation in forecasting
Random splits leak future data into training; walk-forward validation rolls the origin ahead, testing only on later observations. Awareness of temporal leakage. Claiming random splits work for time-series.
How would you probabilistically forecast 40 stories using throughput data?
Tests probabilistic forecasting literacy using historical throughput. Good answers gather 8–12 periods of throughput, run Monte Carlo resampling, and present percentile delivery curves (e.g., 50th/85th/95th).
How is work selected and forecasted for the Sprint Backlog?
Tests empirical forecasting. Outline: the team selects from the ordered Product Backlog using observed experience and expertise to create one valuable Increment. Red flag: treating the forecast as a hard commitment or citing velocity as a required input.
Transform a Time Series for a Gradient Boosting Model
Tests your ability to convert a sequential problem into a tabular one. A great answer covers creating lagged/rolling features and time-based features (e.g., day of week), and crucially, specifies a time-aware validation split.
Describe two methods for generating prediction intervals
This tests your grasp of uncertainty quantification. A great answer contrasts an analytical method (assuming normal errors, using multipliers like 1.96 for 95%) with a simulation method (bootstrapping residuals).
Which model for forecasting with seasonality and trend?
This tests your knowledge of classical time series models. A good answer names Holt-Winters, explaining its level, trend, and seasonal components. It also discusses choosing between additive and multiplicative methods. A red flag is jumping to complex models.
Explain time series stationarity and how to achieve it
Tests your grasp of core time series assumptions. Define stationarity (constant mean/variance over time), explain why models need it for stable predictions, and name methods to test and achieve it. A red flag is just saying the data looks 'flat'.
How would you measure a sales forecast model's accuracy?
Tests if you can link statistical metrics to business outcomes. Define MAE (average error) and RMSE (penalizes large errors). Choose RMSE when large misses are costly (e.g., stock-outs), MAE otherwise. A red flag is reciting formulas without business context.
How would you create a probabilistic forecast for 40 stories?
This tests your ability to use statistical methods for forecasting. A great answer explains how to use historical throughput in a Monte Carlo simulation to generate a probability distribution of completion dates, not a single point estimate.
SARIMA vs. LightGBM for Forecasting with External Variables
Tests your grasp of practical trade-offs in model selection. A strong answer contrasts SARIMA's interpretability with LightGBM's power for handling many non-linear variables, covering performance and implementation costs.
Get Forecasting bites daily.
Five a day, five minutes, offline. With quizzes so it sticks.
Open testing — you’ll join as an early tester.