Case study 01 / Major Czech media company
An explainable forecasting system replaced a decades-old manual process at a nine-figure media business, improved accuracy, and kept human judgment exactly where it mattered.
Client identity and exact commercial impact withheld under NDA.
The brief
Nine-figure USD revenue · Long-term audience forecasting
60 → 1
Hours of human effort per forecast cycle
+15%
Accuracy across every channel and age group
+52%
Accuracy on forecasts the system stood behind
T+2
Months ahead, aligned to inventory sales
The business problem
Long-term viewership forecasts were built manually by one expert using twenty years of accumulated instinct. Each cycle consumed roughly sixty hours, took weeks to finish, and produced a number nobody else could properly explain or reproduce.
The forecast was not an internal planning curiosity. It sat behind the sales team's pitch. Advertisers bought promised views in specific timeslots. When reality came in below the promise, the client could ask for a partial refund or extended service. Every weak forecast created direct P&L exposure.
Underforecasting also left money on the table. A more reliable number meant inventory could be sold with greater confidence rather than hedged against. Accuracy protected downside and unlocked upside at the same time.
The forecast, made visible
This representative August profile shows the shape the system had to understand. The manual forecast drifts around reality. The model stays close, and visibly asks for help when prime-time uncertainty exceeds its authority.
August forecast window / T+2 months
Selected timeslot
21:00
902k forecast viewers
Model confidence
54% / Low
Human review required
Why the system decided this
Prime-time premiere with no direct historical match
Hover or focus any model point to inspect it
Illustrative values based on the production behavior. Client audience figures are anonymized.
The architecture
The quality came from designing around the business, not from selecting the most fashionable algorithm.
Historical audience behavior for the exact group and timeslot was combined with derived relationships, the future programming schedule, seasonality, and calendar effects. Summer audiences alone could run roughly 30 percent below winter.
Midday behaves differently from prime time. Instead of forcing one model across incompatible patterns, the day was split into time blocks. XGBoost, LightGBM, or ExtraTrees was selected according to the shape of each block.
Known patterns moved through automatically. New shows, unusual holidays, and high-volatility slots were flagged with an explanation and routed to the experienced forecaster. The system automated certainty and escalated ambiguity.
Forecasts ran two months ahead and deliberately skipped the intervening month. Those nearer slots were already sold and managed by shorter-horizon models. The boundary followed how the business actually operated.
Human in the loop
The expert did not disappear. Sixty hours of repetitive estimation became roughly one hour reviewing only the timeslots where experience could improve the result.
01 / Model
Forecast every slot
Generate the number, explanation, and confidence score.
02 / Gate
Check confidence
03 / Human
Review exceptions
Approve or adjust the few slots that truly need judgment.
~98%
Routine forecast automated
~1 hr
Expert review per cycle
Under deadline pressure
Two weeks before delivery, the existing evaluation was found to contain data leakage. The model had effectively seen test information during training, and the reported accuracy was not trustworthy. The work had to be rebuilt from the evaluation methodology up while processing millions of records and selecting the right model family for each time block.
The outcome
Accuracy improved 15 percent across every channel and age group. On forecasts that cleared the confidence gate, accuracy improved by roughly 52 percent. The sales team received a stronger commercial foundation, while P&L exposure from missed promises became easier to control.
Fewer weak promises that could trigger refunds or make-goods.
Inventory could be sold against tighter, more credible audience numbers.
Twenty years of judgment became a review layer rather than a key-man dependency.