An Intelligent Fashion Replenishment System Based on Data Analytics and Expert Judgment

Abstract

Retail stock allocation is crucial but challenging. The authors developed an innovative solution, successfully tested in the context of high-end fashion: collaboration between artificial intelligence and human intuition. Each week, stores are assigned a budget based on current stock levels versus potential sales, and offered to “spend” this budget with an initial data-driven recommendation on which SKU/sizes order and release. Each store manager is then given a time window, so she can modify the proposal while respecting budget constraints; and finally, the artificial intelligence optimally allocates available stock to requests based on the expected likelihood of sale minus cost of logistics, subject to management-defined constraints. Our test showed how this system outperformed the control group of stores, relying on a traditional head office-driven allocation without direct human input. The retailer boosted sales, demand cover, and stock rotation performance: an estimated 1M EUR margin/month positive impact. Moreover, the new system improved store managers morale through non-monetary incentive-driven empowerment.

Appendices

Appendix: Forecasting Model

Category sales $\tilde{Y}_C^{\tilde{w}}$ (aggregate: all sizes, all stores, and items in C, with one-year time horizon) are forecasted through multivariate regression on seasonality $\bar{z}$, average full price p, average markdown m, and units per tickets ratio u. The forecasting procedure consists in these steps:

• 1. historical data is aggregated at the category, year, and week level

• 2. for every category C and week of the year w, weekly seasonality index $\bar{z}^{w}_{C}$ is computed

• 3. a linear model $\tilde{Y}=f(z, p, m, u)$ is fitted for every category

• 4. an estimation of the KPI p, m, and u is given for the following 52 weeks

• 5. one-year horizon sales forecast is computed $\tilde{Y}_C^{\tilde{w}}$ for each category and week by using the estimated models

In the following, details on seasonality index and estimation of the regressors for the future year are provided.

A. Category Seasonality

From the weekly aggregated sales, for each category, we remove the effect of markdowns by estimating the model

$$ Y_{C}^w = \alpha _C + \beta _C md_{C}^w. $$

Then for each week, we calculate normalized sales

$$ \bar{Y}_{C}^{w}= Y_{C}^w- \beta _C md_{C}^w. $$

We then average normalized sales $\bar{Y}_{C}^{w}$ over the years, obtaining $\bar{\bar{Y}}_{C}^{w}$, and we smooth this using the moving average of three weeks:

$$ z^{w}_{C}=\frac{\bar{\bar{Y}}_{C}^{w+1}+3 \bar{\bar{Y}}_{C}^{w}+\bar{\bar{Y}}_{C}^{w-1}}{5}. $$

Finally, we normalize seasonality (every category sums to 52)

$$ \bar{\bar{z}}^{w}_{C}= z^{w}_{C} \times \frac{52}{\sum _w z^{w}_{C} }. $$

B. Prediction of Business Indicators

Future average values of full price p, markdown m, and units per tickets ratio u are estimated by moving averages. As an example, starting from category level average full price for a given week and year, $p_{C}^{w}$, we average price over the years obtaining $\bar{p}_{C}^{w}$. Then, we smooth this using the moving average of three weeks:

$$ \bar{\bar{p}}^{w}_{C}=\frac{\bar{p}_{C}^{w+1}+3 \bar{p}_{C}^{w}+\bar{p}_{C}^{w-1}}{5}. $$

$\bar{\bar{m}}$ and $\bar{\bar{u}}$ are then computed in the same way and then used with in (5) as regressors together with $\bar{\bar{p}}$ and $\bar{\bar{z}}$. It is also possible to manually change these values according to future management strategy.