Demand changes.
So does the horizon.

Imagine a planner reviewing next week's replenishment and a buyer committing to several months of purchases. Both need to anticipate sales, but they have different horizons and different opportunities to update their information. This is the motivating scenario for the study, not a client engagement.

The supplied dataset spans January 2018 to December 2023. The notebook explores eight sales categories, then models N02BE using 26 weeks of previous sales. Excess inventory and shortages motivate the question, but their costs are not measured here.

Observed sales are the prediction target. Without stock availability and unmet-order data, they should not be treated as a complete measure of demand.

1. Explore weekly sales. Inspect trends, category correlations and seasonal patterns.
2. Prepare sequences. All 313 observations remain in the modeling series. The 26-week window produces 229 pre-test sequences and 58 holdout targets. A chronological split reserves 35 pre-test sequences for validation; the scaler is fitted on the initial fitting observations.
3. Fit a recurrent model. Two LSTM layers with 64 and 32 units feed dense layers with dropout. Validation selects seven training epochs, then the evaluation model is refitted on the pre-test sequences.
4. Compare at a shared origin. Recursive LSTM, Prophet, ARIMA(2,1,2), persistence and a 52-week seasonal-naive baseline predict the same Monday dates, without receiving holdout sales. A separate all-history LSTM powers the dashboard projections.

One origin, limited generalisation. The comparison aligns dates and information available to each model, but covers one historical cutoff. ARIMA(2,1,2) is an untuned, nonseasonal baseline. Repeat the comparison across several origins before selecting a model.

Features versus inputs. N02BA and N02BE sales are strongly correlated (Pearson r ≈ 0.949). That suggests testing lagged cross-category signals; it does not show patient substitution. Calendar and category features are explored, but the LSTM receives only past N02BE sales.

Sales, uncertainty and value. Sales can be constrained by stock availability. Point projections do not quantify uncertainty, and the study does not measure inventory savings, service-level gains or shortage costs.

1. Prepare for the observed spring pattern. Review purchasing schedules and supplier capacity before March–May. Check recent sales and procurement lead times before setting a seasonal buffer.
2. Choose a model for the planning horizon. Extend the matched-origin comparison to rolling 1-, 4- and 12-week evaluations. Keep Prophet, seasonal naive and LSTM as candidates; assess bias, underprediction and overprediction as well as average error.
3. Measure value in a replenishment pilot. Combine forecasts with on-hand inventory, expiry dates, service targets, lead times, and shortage and holding costs. Validate prediction interval coverage, then compare stockout days, expiry waste and inventory costs with current practice.

The Streamlit dashboard connects eight-category sales charts, an N02BE monthly heatmap, forecast controls and the recorded benchmark tables. Choose 1–12 weeks, inspect the forecast boundary, and download the projected values as CSV.

It loads a saved all-history TensorFlow model and matching scaler. No API key or retraining is needed. The displayed projections begin after the final observation in December 2023, so they should not be read as forecasts for today.