National demand: history and 12-month forecast
Note the shape. Q1 (Apr–Jun) is the peak and Q4 (Jan–Mar) the trough, every year — the inverse of most Indian FMCG books, because the seasonality here is weather- and festival-driven rather than incentive-driven. There is no fiscal-close loading to model. Teal ticks mark south-west monsoon onset, ochre ticks Diwali, pink ticks Vijaya Dashami.
Category forecast · next three months
The volume–value gap column is the point: LED categories grow in units while value stays flat, so a revenue-trained model under-orders them systematically and in one direction. The forward gap shown here is narrower than its FY24–FY26 peak, because management said pricing stabilised in Q1 FY27 — switch the LED realisation lever to see the regime move.
Rolling-origin backtest
There is no incumbent platform to beat. We found no o9, Blue Yonder, Kinaxis or SAP IBP footprint at Eveready, so the honest baselines are seasonal naive and the current planning number — an uncontested comparison, which is a commercial advantage on this account. Five origins at monthly, embargo set to the 8–14 week import and contract-manufacturing indent lead time.
Scenario levers
Modelled drivers, not sliders on the output.
Driver attribution
Method. Ablation deltas on weighted CRPS — retrain without the block, measure the loss. Not attention weights.
Depot despatch: weekly history and 8-week forecast
Distributor indent mix
Share of indent value by segment at this depot, next four weeks. One distributor places one indent across all three segments against one credit line — so a battery scheme pulls lighting volume forward or crowds it out. Move the scheme lever and watch the mix shift.
Why this panel exists. No pure-lighting vendor can offer it. Forecasting the three segments independently double-counts the total and mis-attributes the cause when a category moves.
Product replenishment · next four weeks
Weeks of cover against the P50, and a suggested order at the service level set on the right. The recommendation moves with the quantile, not the mean.
Rolling-origin backtest · weekly
Why weekly errors are larger. Distributor indents arrive in case and full-truckload steps and are timed to scheme slabs, so a weekly bucket carries irreducible lumpiness a monthly bucket smooths away. The event-window column covers the pre-monsoon flashlight weeks and the Puja sell-in weeks — where the stock-outs that matter actually happen.
Scenario levers
Modelled drivers, not sliders on the output.
Driver attribution
Note the reordering versus the national tab. Distributor order behaviour and DMS secondary offtake dominate at weekly depot grain; price regime, macro and input cost drop to almost nothing. Same model family, different weight vector — which is why the two levels are trained separately and reconciled.
Open negotiation items · October 2026 cycle
Where the statistical baseline and the Kolkata regional view disagree by more than the tolerance band. Every item carries a reason code, an evidence read from the model, and an audit trail. Click a row to open it.
Model evidence for the selected item
What the forecast engine can say about the disagreement — the same numbers both sides are arguing over, computed once.
Version history
Every published forecast version, with author, role and the change it made. Click to inspect; the Supply Chain Head can restore.
Roles and permissions
Why this matters more than the model. The commonest cause of failure in forecasting programmes is not model error — it is that the forecast becomes a negotiated number contaminated by targets. Separating forecast, plan and target in the data model, logging every override with an author and a reason, and reporting Forecast Value Add per role is what keeps the statistical number honest.
Ask the forecast in plain English
Questions about volumes, growth, the volume–value gap, cover, orders, accuracy, drivers and the consensus state are answered locally from the model, instantly and deterministically. Anything the local grammar does not recognise is passed to Claude with a digest of the same numbers.
How this works in production. The English layer is a thin tool-calling wrapper over the same forecast store the console reads: intent and entity resolution against the category, product, depot and calendar dimensions, then a typed query. It never invents a number — every answer cites the level, the version and the horizon it came from, which is what makes it safe to put in front of a sales or finance user.
Recent questions
Shared across everyone with the console open. Click to re-run.