Forecasting fashion retail sales in Brazil: a case study before, during and after COVID-19 Open Access

The tertiary sector is a cornerstone in Brazil’s economy, as it provides goods and services directly to end consumers. In 2021, according to The Brazilian Institute of Geography and Statistics (IBGE, 2022) it accounted for 57.76% of the country’s gross domestic product (GDP). Among the industries within this sector, the retail industry is the largest employer in Brazil representing 19.40% of formal job contracts in the country in 2021, as reported by The Ministry of Labor (GDRAIS, 2022).

However, the COVID-19 pandemic and the consequent lockdowns imposed by the government have had a significant impact on the retail industry (Schleper, Gold, Trautrims, & Baldock, 2021). In the second quarter of 2020, retail revenue fell by 14.91% compared to the same period in 2019, as reported by IBGE (2022). Nevertheless, despite traditional sales taking a downturn, e-commerce was able to capture a significant portion of customers who chose to isolate themselves or were left unattended by closed stores (Cramer-Flood, 2020; Mathradas, 2020). This shift in consumer behavior resulted in a 29% increase in the number of Brazilian online shoppers in 2020, followed by a 10% increase in 2021, while profits rose by a record amount of 47% in the first semester of 2020, as reported by Webshoppers (|Ebit|Nielsen, 2020a, 2022).

With the growth of e-commerce, companies have intensified the collection of customer-related data, including purchase histories, preferences, and behavioral patterns. This evolution has increased the strategic importance of data quality and processing because the effectiveness of forecasting models depends on the availability of structured and reliable data. Zhang and Wu (2025) discussed the importance of big data analytics in helping retailers transform customer behavior into actionable insights for more accurate forecasting and decision-making. As highlighted by Queiroz, Ivanov, Dolgui, and Fosso Wamba (2022), the growing complexity of digital operations has driven the demand for business intelligence tools that are capable of supporting decision-making. In this context, forecasting models based on traditional statistical methods and machine learning techniques have become essential for analyzing sales trends and predicting future demand. Seyedan and Mafakheri (2020) pointed out that forecasting models assist businesses in understanding current market behavior and preparing for future demand fluctuations. Several prominent studies (Aras, Deveci Kocakoç, & Polat, 2017; Yu, Tian, & Tao, 2022; Loureiro, Miguéis, & Silva, 2018; Ma & Fildes, 2021; Ramos, Santos, & Rebelo, 2015) have explored forecasting models for retail store sales prediction, including traditional statistical, artificial neural networks, hybrid, and state-space models.

Inaccurate forecasts can potentially trigger stock shortages or excessive inventories, resulting in financial losses for businesses (Nunnari & Nunnari, 2017). This issue is particularly critical in consumer-centric industries, such as fashion, where precise forecasting is crucial (Beheshti-Kashi, Karimi, Thoben, Lütjen, & Teucke, 2015). Precise forecasting of diverse economic factors is crucial in business and economics. This includes predicting Past Economic Performance, Current Global Conditions, Current Industry Conditions, Rate of Inflation, Internal Organizational Changes, Marketing Efforts, Seasonal Demands, etc. (Krishna, Aich, & Hegde, 2018). The development of effective sales forecasting models that produce accurate and robust results is essential for formulating effective strategies that play a crucial role in identifying market changes and understanding consumer behavior, thus enabling traders to formulate market strategies that adapt to evolving trends. In recent years, several approaches based on traditional statistical models and machine learning techniques have been applied in the retail sector, where companies increasingly integrate data modeling into decision-making processes driven by demand forecasting (Loureiro et al., 2018; Nenni, Giustiniano, & Pirolo, 2013). These approaches incorporate pertinent information such as promotional periods and calendar events into their forecasting structures. The adoption of these models stems from the recognition of the profound impact that highly accurate forecasts can have on various operational facets, including inventory and pricing management, production planning, and sales strategies (Fildes, Ma, & Kolassa, 2022).

Fashion retailers often have to deal with volatile sales (Choi, Hui, & Yu, 2014), because the industry faces rapidly changing fashion trends and dynamic markets, making forecasting in this industry a notoriously difficult task (Beheshti-Kashi et al., 2015; Ren, Chan, & Siqin, 2020). Consequently, accurate sales predictions are highly valued by the fashion industry because of the complexity involved in their generation. In this context, a crucial question for retail scholars and managers arises: Can machine learning forecasting models capture potential changes and do certain models outperform others in forecasts under specific scenarios?

To offer a response, the main objective of this study is to compare the performance of several approaches in modeling the daily e-commerce sales of a Brazilian fashion retailer, in full and off-price periods, before, during and after COVID-19. The aim is to make short-term forecasts that will support the commercial team in defining the company’s discount policy regarding when and how much price should be reduced. These forecasts can act as Key Performance Indicators (KPIs) to facilitate informed decision-making. The methodology employs Autoregressive Integrated Moving Average (ARIMA) and Neural Network AutoRegressive (NNAR) models to analyze the e-commerce sales series of a Brazilian fashion retailer. The forecast accuracy was evaluated using Root Mean Square Error (RMSE), Symmetric Mean Absolute Percentage Error (SMAPE), and Total Absolute Percentage Error (TAPE) metrics. The results show that NNAR models exhibit lower forecast errors in both the full- and off-price periods, whereas Seasonal ARIMA performs well in the off-price phase. In general, NNAR is more appropriate for capturing the sales dynamics. Finally, during the fourth collection launched amid COVID-19 restrictions, sales behavior in the full-price and off-price periods was similar, in contrast to previous collections, where sales during the off-price stage were more volatile. These findings contribute to the understanding of model performance in emerging markets and offer practical insights for fashion retailers to manage short-term pricing and inventory strategies. This paper is organized as follows: Section 2 reviews the relevant literature; Section 3 presents the data and methodology; Section 4 details the empirical results; Section 5 discusses managerial implications; and Section 6 concludes with limitations and suggestions for future research.

E-commerce, often referred to as e-commerce, entails the exchange of goods or services over the Internet (Shettar, 2023). It encompasses a diverse range of data, systems, and tools designed for online buyers and sellers, including aspects such as mobile shopping and secure online payment methods (Shettar, 2023). Currently, despite e-commerce undergoing significant expansion and possessing competitive advantages for organizations, it is a highly competitive business model that faces challenges in retaining customers (Veloso, Magueta, Sousa, & Carvalho, 2020).

The use of this sales channel brings benefits to both companies that employ it for sales and consumers. For the company, the introduction of self-service involves shifting certain internal processes, such as product selection and physical payment, to the end consumer. This results in cost-efficiency gains for the retailer. Additionally, data are generated for each consumer, enabling segmentation and targeted activities with a greater focus on the correct consumer market (Burt & Sparks, 2003; Mathradas, 2020).

The fashion and clothing industry is expansive, encompassing a wide range of retail outlets and catering to diverse audiences, including adults, women, men, children, teenagers, beachwear, and intimate apparel. To succeed in this dynamic market, businesses must consistently offer fresh and appealing products, actively engage with consumers, meet their expectations, and demonstrate adaptability and scalability (Choi et al., 2014). In addition, the seasonal nature of fashion means that sales data are limited, making historical forecasting less reliable (Beheshti-Kashi et al., 2015; Choi et al., 2014). Brands are increasingly facing well-informed customers, and with globalization, access to styles from other countries has raised expectations in the domestic market. This has led stores to continually seek innovations and quicker ways to respond to market demand. In summary, the fashion industry frequently encounters significant challenges in achieving accurate forecasts, owing to its short product life cycles, product variability, and demand uncertainties (Beheshti-Kashi et al., 2015). Examples include online purchases with in-store pickups, express shipping, and scheduled deliveries, some of which are already widespread in sectors such as floriculture.

Amid the COVID-19 pandemic, many retailers have adopted omni-channel retailing strategies to withstand operational disruptions, integrating solutions such as upgraded digital platforms, the active use of social networks, and flexible delivery models such as click-and-collect, rapid shipping, and scheduled dispatches (Zhang, Wu, Huang, & Zhang, 2021).

The online market also adapted to accommodate “new buyers”; about 34% of consumers over the age of 55 began shopping online weekly, and nearly 40% made their first online purchases of clothing, footwear, or accessories during the lockdown (Ecola, Lu, & Rohr, 2020; Gupta, Gaurav, & Panigrahi, 2023).

In Brazil, the COVID-19 pandemic significantly accelerated the adoption of e-commerce, not only in sales volume but also in transforming customer experience, particularly in sectors such as supermarket retail. A national study by Cunha, Lettieri, Cadena, and Pereira (2023) highlighted that delivery logistics is the most affected quality dimension. However, this did not significantly impact overall customer satisfaction, suggesting that other factors such as app usability and product presentation played a more decisive role during this digital transition.

This behavioral shift is further supported by data from the Webshoppers reports (|Ebit|Nielsen, 2020a, b, 2022), which indicate that 7.3 million Brazilians made their first online purchase in the first half of 2020, near matching the entire total for 2019. Moreover, 95% of these new consumers expressed an intention to continue shopping online after the pandemic. Supporting this trend, the MCC-ENET index (2022) recorded more than 55% online revenues and a nearly 54% increase in purchase volume in December 2020 compared to the same period in 2019.

In recent years, several statistical models have been proposed to analyze trends in retail sales (Chen & Ou, 2011; Gupta et al., 2023; Lalou, Ponis, & Efthymiou, 2020). Traditional time-series approaches, such as Autoregressive Integrated Moving Average (ARIMA), continue to be widely used because of their ease of interpretation and implementation, particularly in resource-constrained environments (Gupta et al., 2023). Simultaneously, the integration of data analytics and statistical programming has gained prominence as companies seek to handle increasingly complex and volatile demand patterns. For example, Lalou et al. (2020) demonstrated the use of simple forecasting techniques, such as moving averages, in a 3PL apparel distribution context, showing that low-complexity models, when supported by structured data processing, can significantly enhance inventory management and reduce operational workload in distribution centers.

More recent developments include univariate machine learning forecast models that leverage historical sales data and/or explanatory variables, such as promotional dates, calendar events, weather conditions, and other relevant factors, to account for seasonal trends and behavioral patterns. Some studies suggest that incorporating exogenous variables enhances the accuracy of forecasts in terms of the Stock Keeping Unit (SKU) level (Kourentzes & Petropoulos, 2016; Ma, Fildes, & Huang, 2016). Wolpert and Macready (1997) emphasized that no single forecasting method can consistently exhibit superior performance across all possible datasets, highlighting the importance of selecting models based on the specific characteristics of each problem. This suggests that the dominance of a single method over the others for all products and future time periods is improbable.

Among the univariate models without explanatory variables in their structures, the ARIMA, Exponential Smoothing family models (ES), and state-space model are noteworthy in the retail sector. On the other hand, Autoregressive Integrated Moving Average models with exogenous variables (ARIMAX) and Seasonal ARIMAX (SARIMAX) models allow the inclusion of external variables, as well as the modeling of trends and seasonal patterns (Koopman & Durbin, 2012). Finally, recent research has shown improvements in forecasting accuracy when using nonlinear models over linear regressions such as Bayesian P-splines and recurrent neural networks (Bianchi, Maiorino, Kampffmeyer, Rizzi, & Jenssen, 2017; Brezger & Lang, 2006; Terui & van Dijk, 2002).

In the fashion field, some recent studies have found relevant results when using traditional forecasting models. In Ren et al. (2020) analyzed the daily and weekly sales of a real-world case in which a Chinese company used ARIMA models. The forecasts were best for basic products, instead of fashionable items, since the sales of the former were more stable over time. In Rubio, Gutiérrez-Rodríguez, and Forero (2021) compared the performance of Exponential Smoothing (ES) and ARIMA models to predict the values of a real profit measure for the entire fashion retail sector in Colombia, where economic and social changes have impacted the industry, resulting in a notable decline in its actual profit. The historical data reported over a period of 24 years were considered. The researchers found that ARIMA was more accurate in its forecasts than the exponential smoothing models tested based on accuracy metrics, including simple exponential smoothing (SES) and double exponential smoothing (DES).

Güven and Şimşir (2020) also described another attempt to forecast garment sales. This study aims to achieve accurate predictions using two artificial intelligence models, Artificial Neural Network (ANN) and Support vector machines (SVM), based on historical sales data and information about products individually, such as color and directed gender. By comparing the accuracy of the models through the RMSE of both methods, ANN showed superior performance to SVM in seven out of ten models, excluding product color as a covariate, while both models performed similarly when color information was incorporated. To the best of our knowledge, there is a scarcity of studies that compare the performance of linear models to ANN in the post COVID-19 economy, where new consumption patterns may have emerged. In this context, according to Au, Choi, and Yu (2008), models with network structures provide more accurate short-term sales forecasts, particularly in environments characterized by low demand uncertainty and weak seasonal trends, compared to traditional statistical models such as Seasonal Autoregressive Integrated Moving Average (SARIMA).

In this study, the ARIMA and NNAR models are used to study the behavior of the e-commerce sales series of a Brazilian fashion retailer and make short-term forecasts. The data were divided into a training set and a test set. The first set was used to estimate the model parameters, and the second set was used to evaluate the accuracy of the models on data not used during model training. RMSE, SMAPE, and TAPE were the metrics used in this study to assess the accuracy of the forecasts.

The ARIMA models allow modeling of non-stationary homogenous processes, denoted by ARIMA(p,d,q), where $d$ indicates the number of differences needed for the series to become stationary, $p$ is the total number of lagged terms of the series that will be considered in the Autoregressive (AR) part, and $q$ is the total number of lagged error terms that will be considered in the Moving Average (MA) part (Morettin & Toloi, 2018). It is worth noting that a time series is considered stationary when the distribution of the data is independent of time, that is, the mean remains constant and the covariance between any two values depends only on the time difference between them (Hyndman & Athanasopoulos, 2018). Let $Yt$ be the value of the series at time $t$⁠, The ARIMA model can be represented as

where $Wt$ is the time series after $d$ differences; $ϕ1,…,ϕp$ are the autoregressive (AR) parameters to be estimated, which captures the linear dependence of the current value on its past $p$ lags; and $θ1,…,θq$ are the moving average (MA) coefficients, which model the linear dependence of the current value on the past $q$ error terms; and $et$ is the error term at time $t$⁠, which generally follows a normal distribution with a mean of zero and variance $σ2$⁠. Finally, in the SARIMA models, the seasonal AR and MA terms were added to the ARIMA structure.

Artificial neural networks are forecasting methods based on simple mathematical models of the brain, which enable the modeling of complex nonlinear relationships between the response variable and its predictors. The simplest networks have no hidden layers and are equivalent to linear regressions, where forecasts are obtained using a linear combination of inputs. NNAR models, such as ARIMA models, incorporate lagged terms of the series as inputs into this structure. The NNAR (p,k) model can be written as:

where $Yt−i,i=1,…,p$ are the lagged values used as inputs and $f(.;w)$ is a neural network activation function parameterized by weights $w$ and with $n$ hidden nodes in a single layer, and $et$ is an error term with mean zero and variance $σe2.$ Generally, a sigmoid function is used as the activation function. Seasonal data are denoted by NNAR (p, P, k), where p represents lagged terms, P denotes terms with lagged value multiples of the seasonal period, and k is the number of nodes in a hidden layer. This structure allows the model to capture nonlinear and seasonal patterns in a time series using a feedforward neural network with a hidden layer containing k neurons. The weights of the lagged terms are selected using a “learning algorithm” that minimizes a “cost function” such as the Mean Squared Error (MSE), which is calculated by averaging the squared differences between the predicted and actual values of each data point in the dataset (Hyndman & Athanasopoulos, 2018).

The metrics used to measure the performance of the models are given by:

where $Yt$ and $Yˆt,$ represent the observed value of the series and its forecasted value at time $t$⁠. It is worth noting that the TAPE criterion compares the total predicted and observed weekly sales, aligned with the company’s performance-monitoring approach.

The data used in this study were provided by a Brazilian fashion retailer that stored transaction information in the SQL Server Management Studio (SSMS) database. The well-established retail company, which provided the data, was founded in the 1980s, and it exceeded the typical lifespan of most Brazilian companies, which, on average, demonstrated a 20% survival rate at the 12-year mark (Ehrl, 2021). More recently, in the second quarter of 2018, the firm started an online e-commerce branch. Its operational processes have reached a significant level of maturity, allowing reliable data collection and analysis. The company in question is a large-scale enterprise specializing in contemporary and sophisticated women’s apparel, including blouses, shirts, pants, skirts, and dresses. Its target market comprises modern urban women, primarily from the middle to upper classes, offering versatile and elegant garments suitable for a variety of occasions. While the brand’s pricing policy allows accessibility to a less affluent audience, many items are still considered expensive without discounts. The store location also reinforces this positioning, with its flagship store in an upscale neighborhood in São Paulo’s southern zone and others distributed across prestigious areas throughout the country. This study uses sales data from eight collections spanning July 2018 to August 2022: Collection 1 (2018.07–2019.03) through Collection 8 (2022.02–2022.08).

Table 1 presents the total daily sales of the retailer’s e-commerce channel in reais (R$), grouped by clothing collection, from July 2018 to August 2022.

The company produces two clothing collections annually. Once a collection is released to the public, its price remains undiscounted for an average period of 203 days for winter collections and 162 days for summer collections. After the full-price period, the collection enters an off-price phase, in which prices are cut progressively until the products are sold completely. This pricing method has been found to increase sales, although the promotional effect on profits diminishes over time (Li, Yada, & Zennyo, 2021). The active discount policy periods of the collections are asynchronous and disconnected, which explains the clustering approach adopted in this study, as this study aims to understand the effect of discount periods on sales. In addition to the off-price stage of the collections provided to all customers, the discount series considers other discounts, including coupons obtained individually by clients through the company’s benefit programs.

It is important to note that items sold by the retailer primarily target female audiences. The items sold are shirts, pants, skirts, dresses, and jackets. In addition, the retailer offers other clothing and accessories. The number of products released in each collection as well as the composition of the product groups varied significantly. The classification of each item becomes intricate as continual innovations in product design introduce additional features, thus challenging the applicability of pre-established item group definitions.

The sales data encompass eight collections, including the full price stage as a whole and the first two months of the off-price stage for each collection. This analysis period covers approximately 75.4% of the recorded sales for each collection. A descriptive analysis of each collection’s sales is presented in Table 2, including minimum (Min), maximum (Max), mean, and standard deviation (Std) values in reais (R$) for both full- and off-price stages across the entire analysis period.

Note that in all collections during both price stages, there were days when no sales were recorded, except for the eighth collection. During the off-price period of the eighth collection, the lowest daily sales recorded was R$ 849. It is important to highlight that, generally, collections experience days without sales for approximately 25% of the exhibition period, with a slightly higher occurrence in winter collections. The average sales for the eighth collection, launched in February of 2022, was R$ 12,119, making it the second collection with the highest average sales recorded during the full price period.

The fourth collection, which was available to the public between February and August 2020, when stores in Brazil were closed to reduce the spread of COVID-19, had the highest average sales and the highest daily sales recorded during the full price period. Compared with the second collection, which was sold during a similar time in the previous year, the average sales increased by 67.5%. It is also worth noting that the fourth collection exhibited high sales volatility (the highest standard deviation) and maximum daily sales during the full price period among all collections. This pattern could be attributed to the surge in online sales via the e-commerce channel in 2020, as more customers stayed home due to the pandemic, resulting in more volatile buying behavior and higher peak sales for this collection during its full-price stage. On the other hand, in the off-price period, the fourth collection showed an increase in volatility and mean sales, which can be observed for all other collections as well. However, this increase in volatility was the smallest when compared with the other collections, resulting in the fourth collection being the least volatile in the off-price stage, presenting consistent sales and low peak values.

Furthermore, collections 5 to 8 re-established the sales behavior recorded before the pandemic, with low sales and variability during the full-price period and high volatile sales during the off-price stage. These collections were sold during a period when COVID-19 restrictions were eased, and commerce returned to its pre-pandemic mode of operation.

It is worth highlighting the weak sales performance of the sixth collection, launched at the beginning of 2021. It had the lowest sales and little variability in both the full- and off-price periods. At the same time, retail in Brazil regained traction after the first hit it had taken in 2020, presenting a 47.8% increase during the second quarter of 2021 (IBGE, 2022). This contrast suggests that fashion retail does not necessarily follow broader retail trends, which exemplifies the difficulty of forecasting fashion sales mentioned in the literature (Au et al., 2008; Fildes et al., 2022; Ren et al., 2020). As pointed out by Krishna et al. (2018) in the realm of business and economics, diverse economic factors can impact forecasting, as country environmental conditions such as unemployment rate in Brazil that reached 14.8% during this period (IBGE, 2022), which may correlate with the low sales observed in the collection. In a scenario where the monthly budget is limited, the population may have needed to exclude low-priority items from their expenses.

The sales in reais (R$) during the study period are presented in Figure 1. The dotted red lines indicate the beginning of each collection.

It is observed that, in all collections, sales increase during the off-price period, presenting the highest daily sales during this period. However, the fourth collection is the only one that does not exhibit this behavior. Sales recorded during the off-price stage of this collection showed less variability and significantly lower peak values than the other collections.

It is worth mentioning that the fifth collection was delayed due to COVID-19 related issues, which led to a time window not being considered in the study since it surpassed the two months of off-price stage analysis of the prior collection.

To identify seasonal aspects of sales, the spectral densities of the collections’ time series were measured and observed using a periodogram. The collections share a repeating pattern that can be observed every week. Revenue peaks normally occur on Mondays; when compared with the mean sales of the other weekdays, Mondays surpassed them by 181.0% in sales. This pattern is a consequence of the invoicing policy of the company, which bills most of its Sunday sales on the following working day, so that the staff can deal with eventual invoicing problems more rapidly.

Daily sales during both the full- and off-price periods for each collection were modeled using the SARIMA and NNAR models. A dummy variable is included in the structure of the models to indicate the period in which sales occurred. The model parameters defined in equations (1) and (2) were estimated using the maximum likelihood and least squares method, as proposed by Hyndman and Athanasopoulos (2018). Sales recorded during the last week of each stage were used for the evaluation. The performance of the models in forecasting short-term sales values was assessed using the RMSE, SMAPE, and TAPE criteria as defined in Equations (3), (4), and (5). The models were implemented using R programming language with the packages TSA (version 1.3) and forecast (version 8.21.1).

It is important to mention that for most collections, daily sales exhibited a seasonal pattern occurring every seven days. In other words, the sales recorded at time $t$ show a stronger correlation with the sales recorded at time $t−7i$⁠, where $i=1,2,…$⁠. Thus, to forecast sales for a Monday, it is necessary to consider the information on sales recorded on the previous Monday. Therefore, the SARIMA(p,d,q)(P,D,Q)₇ models, with a seasonal period of seven days, were considered in this study. Table 3 presents the SARIMA models selected for each collection, based on a grid search conducted using different combinations of parameters. The selection was determined using the Akaike information criterion (AIC), which assesses how well a model fits the data (Vrieze, 2012).

Notably, the SARIMA model used for forecasting the sales of the fourth collection, the collection most impacted by the COVID lockdowns, indicates that there is a strong dependence on the seasonal component (P = 4), suggesting a more stable behavior, allowing the model to benefit from information from four weeks prior to the predicted period. The SARIMA models of collections 2, 3, 5, 6, and 7 indicate that a history of up to two weeks is relevant for the forecasts. In these collections, sales volatility increased significantly during the off-price period.

Table 3 indicates that daily forecasts during the full-price period were weak for the first seven collections, with SMAPE values of around 0.9, meaning that the average forecasting error corresponds to half the scale of the actual and predicted values. This reflects the difficulty in accurately forecasting daily sales in the fashion e-commerce industry. Notably, SMAPE is sensitive to values close to zero, and the collection of the company in question goes through days without sales for approximately 25% of the exposure period. However, the traditional models demonstrated good performance in forecasting the total weekly sales. The TAPE values varied around 18% of the total weekly sales volume, indicating good predictive accuracy, particularly given that the TAPE criterion assesses performance based on aggregated weekly values, consistent with the company’s decision-making policies.

Finally, it is worth noting that during the off-price period, the SARIMA models have difficulty predicting sales due to the high variability of sales and the lack of a clear purchasing behavior pattern. Nonetheless, it is worth highlighting the high forecasting accuracy during the final week of sales in Collection 4 with a TAPE value of 0.06. This indicates that on average, the absolute difference between the predicted and observed sales accounted for only 6% of the total actual weekly sales. During this period, sales remained relatively stable and exhibited a behavior similar to that observed during the full price period. Thus, the use of a four-week sales history proved useful in predicting sales during this period. The assumptions of the SARIMA models were verified, and the residuals had a mean close to zero, showed no significant autocorrelation (indicating independence), and followed an approximately normal distribution. However, it is worth noting that residual variability is higher during the sales period, a behavior expected due to the high volatility and greater unpredictability of sales in this phase, as the company runs marketing campaigns based on its own strategic guidelines. The residual analyses are presented in the Supplementary Material (Figure 3).

NNAR models were considered in this study to evaluate their performance in forecasting the sales values. In the NNAR (p,P,k) models, p are lagged values of the series, P are terms with lagged value multiples of the seasonal period, and k are nodes in the hidden layer. Note that all the models use a history of at least seven values to predict the series values. Table 4 presents the results. The selected models minimized the AIC criterion.

It is observed that most NNAR models present better forecasting performance in both the full- and off-price stages when compared with the SARIMA models. Notably, the sales forecasts during off-price stages were more accurate for collections 1, 4, and 8, whose NNAR models presented the smallest number of lagged values in the series. During these sales periods, daily sales were generally more consistent and less variability.

To illustrate the performance of the models’ forecasts, Figure 2 presents the forecasts for the two stages of sales of the fourth collection. It should be noted that the forecasts made by both models follow sales behavior. The forecasting performance of the models for the other collections is presented in the Supplementary Material (Figure 4).

Forecast errors in the Full-price and Off-price periods were smaller when using NNAR models based on neural networks. The SARIMA models also performed well, particularly during off-price periods. Thus, in the context of this company’s data, NNAR models proved to be more suitable as they more appropriately captured the underlying behavior of the data across both pricing stages. It is worth noting that exponential smoothing models were tested during one collection sales for model selection but were excluded from the main analysis due to poor forecasting performance. These findings should be further explored in future studies. These findings are in agreement with a study by Au et al. (2008) in which researchers were able to achieve more accurate fashion product demand forecasts from an ANN, in comparison to a traditional SARIMA model.

It is worth noting that the SARIMA models indicate that daily sales increase by 20% to 30% during the off-price period. This increase is observed across all collections, with the smallest increase occurring in collections 4, 6, and 8, where daily sales increased between 10 and 20% during the discount period.

The behavior of daily sales during the fourth collection, launched during the period in which restriction policies were implemented by the Brazilian government to reduce the contagion of COVID-19, changed when compared to the daily sales observed in the previous collections. During this period, the daily sales registered during the full- and off-price periods exhibited similar behavior. After the restrictions eased, sales returned to their pre-pandemic behavior, presenting low volatility during full-price periods and high volatility during the discount period. Notably, the sales performance of the sixth collection, launched in early 2021, was the lowest. During this period, the unemployment rate in Brazil reached a historically high level, and retail trade showed a significant increase, suggesting that fashion sector sales may not follow the behavior of the retail sector in general.

In summary, the main contribution of this study is the proposal of a methodology that can be implemented in the fashion retail sector to analyze the behavior of sales over time to make short-term forecasts to support decision-making. The impact of the discount period stands out, in which sales increase significantly and purchase behavior is more volatile. The relevance of the tested time-series models for forecasting retail demand, specifically regarding the use of forecasts in managing discount policies, is a basis for future researchers and companies that want to explore similar methods.

From a managerial perspective, the proposed approach offers valuable contributions to the retail sector by providing short-term sales forecasts that support data-driven decision making, particularly in pricing strategies, promotional planning, and inventory management. In a post-pandemic context, where retailers must reassess the operational structures established during the crisis, accurate forecasting becomes even more critical (Cunha et al., 2023). The forecasting models presented in this study enable fashion retailers to anticipate demand across different stages of the product life cycle (full price and off-price), facilitating more efficient resource allocation and reducing last-mile delivery inefficiencies. To support practical applications and encourage replication, the R codes used to implement the SARIMA and NNAR models are provided as supplementary materials.

Although ARIMA is one of the most traditional linear forecasting models, recent studies have suggested that ANNs may offer a promising alternative. Comparisons between the ARIMA and ANN models yield mixed conclusions regarding forecasting performance superiority, depending on factors such as data characteristics, volatility, and the inclusion of exogenous variables (Zhang, 2003).

The findings of this study align with those of previous research showing that neural networks often outperform traditional time-series models in short-term retail forecasting. ARIMA models have significantly improved retail sales forecasting by capturing market dynamics, extracting features, and generating accurate predictions (Qadrini, Asrirawan, Mahmudah, Fahmuddin, & Amri, 2020). Studies comparing ARIMA and Long Short-Term Memory (LSTM) models in retail settings indicate that both are effective; LSTM excels at capturing complex temporal patterns, while ARIMA relies on linear dependencies in historical data (Hameed, 2023). According to Au et al. (2008), models with neural network structures outperform Seasonal ARIMA models for products with lower demand uncertainty (e.g. basic t-shirts), whereas SARIMA performs better for products with higher demand variability.

A limitation of this study is that the analyses conducted within the context of sales in the fashion retail sector cannot be generalized, as this is a case study. In future research, explanatory variables should be included in the model to assess their impact on sales forecasting. Additionally, more time-series models will be implemented and compared, including Long Short-Term Memory (LSTM), a type of recurrent artificial neural network capable of capturing complex temporal dependencies. Generalized Autoregressive Conditional Heteroskedasticity (GARCH) models, which are designed to model time-varying volatility commonly applied in economic and financial data, will also be considered to analyze the impact of volatility on forecasts, particularly given the unexpectedly high variability observed during the SALE stage compared to the more stable full price phase. The combination of SARIMA and GARCH models was also tested.

From a generalizability perspective, although the methodology applied in this study, based on SARIMA and NNAR models, can in principle be implemented in other companies within the Brazilian retail sector, the single company focus naturally limits the generalizability of the results. While the COVID-19 restrictions imposed in Brazil broadly affected the fashion retail industry, the intensity and duration of this impact likely varied across companies. In addition, fashion retail sales are inherently volatile and influenced not only by macroeconomic factors but also by rapid shifts in consumer preferences and trends (Nenni et al., 2013), which makes the assessment and comparison of forecasting model performance across firms more complex. It is also important to recognize that although the sector experienced recovery after the pandemic, internal dynamics, such as managerial strategies, supply chain resilience, and customer profiles, can result in heterogeneous outcomes. Therefore, sector-wide inferences should be made with caution, and the findings presented here should primarily be interpreted within the context of the company studied while offering a reference for future comparative research.

The supplementary material for this article can be found online.