A forecast is a relay, not a hero
The most common way forecasting fails in an organisation is not a bad model - it is a broken handoff. The analyst never ships a baseline, so nobody knows what "good" is. The model works in a notebook but no one productionises it. The forecast is beautiful but never reaches the planning tool. Or nobody watches it drift. Treat forecasting as a relay of five roles, each passing one clear artifact, and you can see exactly where yours is dropping the baton.
The relay and its handoffs 10 min live
Read it left to right. Each box is a role; the light box beneath is the one artifact it hands to the next runner. If an artifact is missing, the relay stalls there - which is exactly where to look when a forecast is not landing.
LiveThe first three runners - profile, model, serve4 min▶
- Data analyst - profiles the history. Plots the series, decomposes it into trend and seasonality, spots outliers, and ships the baseline. The handoff is a baseline plus the plots that explain the series. Maps to builder sessions b1-b2.
- Data scientist - models and backtests. Climbs the method ladder (ETS, ARIMA, Prophet, ML), then selects the winner on a backtest against the baseline. The handoff is a backtested, chosen model with its accuracy numbers. Maps to b3-b8.
- ML engineer / DevOps - productionises. Turns the notebook model into something that runs on a schedule: retrain cadence, batch scoring, and the forecast API. The handoff is a served forecast other systems can call. Maps to b9.
LiveThe last two runners - surface, and monitor4 min▶
- Engineering / product - surfaces the forecast. A forecast nobody sees changes nothing. This role puts the number where the decision happens: a dashboard, or the field inside the planning or inventory tool. The handoff is the forecast visible at the point of decision. Maps to b9.
- Everyone, led by DS / analytics - monitors. Once live, the forecast is tracked against what actually happened. Errors are logged, drift is watched, and a rule decides when to retrain. The handoff is error tracking and alerts that close the loop. Maps to b10.
The two ends fail most often. Teams over-invest in the middle (the model) and under-invest in the ends - shipping without a baseline, or launching with no monitoring. A model that is never watched will drift into being wrong, silently, right when the business trusts it most.
The RACI grid 6 min live
Same five tasks, now as ownership. A = accountable (the one name on the hook), R = also does the work, C = consulted. Notice the two lead roles: the data scientist owns the modelling and the closed loop; monitoring is a shared job the whole team runs.
Self-studySmall teams wear several hats2 min read▶
These are five roles, not necessarily five people. In a small org one analyst-plus-DS may cover the first two, an engineer the next two, and monitoring is a shared habit.
- Do not skip a role by skipping a person. If nobody owns monitoring, monitoring does not happen - the task still needs an accountable name.
- The accountable name matters most. One clear "A" per task prevents the "I thought you had it" failure.
- Buy-vs-build shifts the middle. A vendor forecast product can cover serving, but you still own profiling, the baseline, surfacing, and monitoring.
Find the missing runner ★ 6 min · think it through
A forecast in your org "isn't working." Walk the relay to locate the break instead of blaming the model.
Ask for the baseline. If no one can tell you the seasonal-naive number the model beats, the analyst handoff is missing - start there.
Ask where the forecast lives. If it only exists in a notebook, the ML-engineer handoff (the served forecast) never happened.
Ask who sees it. If the number never reaches the planning tool or a dashboard, the engineering/product handoff is the break.
Ask who watches it. If nobody tracks forecast error over time, the model may already be drifting and no one knows. That is the monitoring gap.
Try it yourself - this week ◐ 20-30 min total
- Take one live or planned forecast in your org and map it onto the five roles - profile, model, serve, surface, monitor.
- Write the name (or "nobody") accountable for each of the five tasks.
- Circle the gap: which handoff has no clear owner or no artifact today?
- Write one sentence on how you would close that gap - hire, reassign, or a vendor for the middle.
- Optional: check that a baseline number exists at all. If it does not, that is your first fix.
Sources covered (cited)
This session reflects standard forecasting-team practice, mapped onto the builder track that runs it in code. It covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · What artifact does the data analyst hand off?
The analyst profiles the series (trend, seasonality, outliers) and ships the baseline. That baseline is the yardstick every later role needs.
2 · Who is accountable for modelling and selecting on a backtest?
The data scientist climbs the method ladder and picks the winner on a backtest against the baseline - the "A" on the model + backtest row.
3 · A forecast "isn't working." What is the leader's first move?
Nine times in ten the problem is a broken handoff, not the model. Name the runner and the missing artifact, then fix that link.