Time Series Forecasting Recruiting

Time series forecasting is the discipline of predicting ordered data: demand, sales, traffic, energy load, and the macro series that steer budgets. Its toolkit spans exponential smoothing, ARIMA models, state space models, and the seasonal adjustment machinery used by official statistics agencies. The field's reference text, Hyndman and Athanasopoulos's Forecasting: Principles and Practice, is written around forecasting as a planning tool, where method choice follows the data and the decision it serves, from judgmental baselines to dynamic regression with ARIMA errors [1][1] Forecasting: Principles and Practice (3rd ed) — OTexts (Hyndman and Athanasopoulos) (accessed 2026-09-28). statsmodels' state space layer shows the statistical machinery underneath: SARIMAX models, Kalman filtering, and residual diagnostics such as Ljung-Box and heteroskedasticity tests [2][2] Time Series Analysis by State Space Methods — statsmodels Documentation (accessed 2026-09-28). Hiring for this craft means finding people who have evaluated forecasts against held-out horizons, not just fitted models.

Challenges in Time Series Forecasting Recruiting

Demand forecasting at scale outgrew hand-tuned models

Organizations forecast thousands of series: one per SKU, per store, per region. fpp3 dedicates a chapter to hierarchical and grouped series, where bottom-up, top-down, and reconciled forecasts must agree across levels, with reconciliation adjusting individual forecasts so that parents and children sum consistently [1][1] Forecasting: Principles and Practice (3rd ed) — OTexts (Hyndman and Athanasopoulos) (accessed 2026-09-28). The industrial consequence is that the scarce hire is not the person who can fit one beautiful model but the one who can run thousands with defensible defaults, monitor them, and intervene where error grows. Demand forecasting seats therefore look less like statistical consulting and more like systems work, which is why automated forecasting systems exist and why CVs describing one competition model say nothing about the job. The interview that finds these people asks about the fleet, not the flagship: how many series, which families per series, how models were retrained, and how forecast degradation was caught before planners felt it.

ARIMA models still win when the data is a measurement, not a business calendar

SARIMAX in statsmodels is seasonal ARIMA with exogenous regressors, estimated by maximum likelihood and run through the Kalman filter, with standard errors and diagnostic tests attached to the results [2][2] Time Series Analysis by State Space Methods — statsmodels Documentation (accessed 2026-09-28). fpp3 gives the modeling logic: stationarity and differencing first, then order selection with information criteria, and the honest comparison between the ARIMA and exponential smoothing families [1][1] Forecasting: Principles and Practice (3rd ed) — OTexts (Hyndman and Athanasopoulos) (accessed 2026-09-28). The craft difference is in who uses the notation. ARIMA models reward people who understand the generating process: what a lag operator means, when differencing destroys signal, why seasonal differencing on retail data is a choice rather than a default. Candidates who can only call fit() on a library can answer none of that, and their forecasts look fine until the data stops cooperating.

State space models carry the cases ARIMA notation cannot express

statsmodels' state space framework generalizes the family: a latent state evolves through a transition matrix, observations load through a design matrix, and SARIMAX, unobserved components, and Holt-Winters exponential smoothing are all one framework estimated through the Kalman filter and smoother [2][2] Time Series Analysis by State Space Methods — statsmodels Documentation (accessed 2026-09-28). That generality buys real temporal modeling capacity: missing observations handled natively, structural breaks and cycles through unobserved components, and a news method that measures the impact of data revisions on estimates [2][2] Time Series Analysis by State Space Methods — statsmodels Documentation (accessed 2026-09-28). Practitioners who have worked in state space models can explain when a local level with a stochastic cycle beats a boxed ARIMA order, which is exactly the judgment that library APIs hide. The framework also carries a higher entry cost: matrix notation, filter initialization, and likelihood work scare off the tutorial crowd, which keeps the experienced population small and genuinely scarce.

Prophet earns its seat on holidays, changepoints and regressors

Prophet models time series as trend, seasonality, and holiday effects, with seasonalities built from Fourier series, weekly defaulting to order 3 and yearly to order 10 [3][3] Seasonality, Holiday Effects, And Regressors — Prophet Documentation (accessed 2026-09-28). Its real strength is calendar work: holiday windows with lower and upper bounds, conditional seasonalities that switch weekly patterns between seasons, and additional regressors such as weather or promotions that enter the linear component [3][3] Seasonality, Holiday Effects, And Regressors — Prophet Documentation (accessed 2026-09-28). The traps are documented as clearly as the features. Fitting monthly data and asking for daily forecasts produces yearly seasonality that is unidentifiable between months and overfits where no observations exist, so forecasts must be made at the monthly frequency [4][4] Non-Daily Data — Prophet Documentation (accessed 2026-09-28). Practitioners who have run Prophet in production know the traps; tutorial users know the defaults.

Time-dependent feature engineering leaks unless the future is known

Every extra regressor must be known for both the history and the future dates, either because it is a calendar fact or because it has been forecast separately [3][3] Seasonality, Holiday Effects, And Regressors — Prophet Documentation (accessed 2026-09-28). The documentation states the failure plainly: when one series is used as a regressor for another, errors in the regressor forecast become errors in the target, and the approach only pays when the regressor is easier to forecast than the target, as in hierarchical settings with a strong top-level signal [3][3] Seasonality, Holiday Effects, And Regressors — Prophet Documentation (accessed 2026-09-28). Time-dependent feature engineering is therefore leak-prone by construction: a rolling average that touches future rows, a holiday flag that ends at the training cutoff, a promotion calendar edited retroactively. The verification question writes itself: how were the future values of every feature obtained at scoring time. A correct answer names the mechanism for each feature; a vague answer means the model has been trained on knowledge it will not have in production.

Seasonal adjustment is a public statistics craft most forecasters never touch

The US Census Bureau's X-13ARIMA-SEATS is the reference seasonal adjustment program: regression models with ARIMA errors, known as regARIMA, ARIMA model-based adjustment via SEATS from the Bank of Spain, the nonparametric X-11 procedure, and diagnostics of adjustment quality and stability, with batch processing of many series [5][5] X-13ARIMA-SEATS Seasonal Adjustment Program — U.S. Census Bureau (accessed 2026-09-28). Official statistics runs on it, from employment to trade. Almost no commercial forecasting hire has used it, and the gap matters in finance and macro-adjacent roles, where seasonally adjusted figures are the input and someone must know what adjustment did to the signal. A forecaster who can explain what regARIMA removes and what remains is a different hire from one who has only ever smoothed. The same gap shows inside banks and energy desks, where adjustment choices feed into every reported rate, and an employer who skips the question inherits numbers whose seasonal story was decided by someone else's defaults.

Backtesting horizons settle econometric modeling claims

Forecasting CVs list models; verification asks for evaluation. Prophet's diagnostics formalize the practice: select cutoff points in history, fit only up to each cutoff, compare forecasts to actuals over the horizon, with rolling performance metrics and a default initial window of three times the horizon [6][6] Diagnostics — Prophet Documentation (accessed 2026-09-28). fpp3 prescribes the same discipline as time series cross-validation and training-set separation, and statsmodels supplies the residual diagnostics, Ljung-Box for serial correlation and heteroskedasticity tests, that tell whether a fitted econometric model is even adequate [1][1] Forecasting: Principles and Practice (3rd ed) — OTexts (Hyndman and Athanasopoulos) (accessed 2026-09-28)[2][2] Time Series Analysis by State Space Methods — statsmodels Documentation (accessed 2026-09-28). The questions that matter: which horizon was held out, what did error look like as a function of horizon distance, and what happened when the candidate's forecast was wrong in production. A model without a backtest is a decoration.