Abstract
This study evaluates and forecasts the usage of the OECD (Organisation for Economic Co-operation and Development) iLibrary database, subscribed to by the Library of the Grand National Assembly of Türkiye (GNAT). To capture dynamic and scale-dependent relationships between usage and external shocks we applied Morlet and Mexican hat wavelet coherence analysis, which are effective tools for continuous wavelet coherence analysis. The results of Morlet wavelet coherence revealed short- and medium-term co-movements driven by external factors, whereas Mexican hat coherence identified more persistent long-term relationships. Subsequently, the regression algorithms were applied on usage data decomposed by maximal overlap discrete wavelet transform (MODWT) and multiresolution analysis (MRA), providing a flexible framework for representing usage data at multiple resolution levels, while preserving temporal alignment and minimizing information loss. The findings revealed that usage patterns exhibited a multi-year synchronization with exogenous factors and regularized linear approaches such as ElasticNet, Lasso, and Ridge regression outperformed non-linear methods. SHapley Additive exPlanations-based interpretation, which quantified the contribution of each predictor of model outputs, showed that short-term fluctuations in usage data were the strongest predictors of future usage. The study demonstrates that e-resource usage should be understood and managed as a multi-scale, dynamically evolving process, and provides a framework that enables libraries to anticipate demand, align collections with institutional cycles, and respond to both short- and long-term changes.
Keywords
Introduction
The serial pricing analyses published annually by Library Journal since the 1960s indicate that, although much has changed in the publishing ecosystem, the costs of information resources have shown a persistent upward trend. This is a long-term and structural problem. For this reason, particularly when preparing for the next subscription renewal period, it is essential to conduct evaluations based on scientific methods and empirical data, in order to manage budgets more effectively in the face of rising costs (Romaine et al., 2025).
By analyzing past usage data with econometric models, it is possible to forecast future electronic resource (e-resource) usage. Econometric models enable the proactive identification of potential issues, and support rational decision-making by employing data-driven insights into collection management. However, traditional econometric models are generally based on assumptions of stationary and linear relationships, which render them inadequate for explaining complex dynamics. Econometric models which assume stationarity may implicitly presume that usage patterns observed in past years—such as steady growth in database usage or predictable seasonal peaks during examination periods—will persist unchanged into the future. However, in real-world settings, library usage is typically non-stationary, exhibiting seasonal fluctuations driven by the academic calendar, sensitivity to sudden shocks such as pandemics, and responsiveness to economic conditions.
Signal processing techniques which enable the analysis of non-stationary time series have been widely used across many fields, including mathematics, finance, and economics. Signal processing methods, such as wavelet transforms, allow the simultaneous analysis of non-stationary time series in both the time and frequency domains, making it possible to examine long-term trends, medium-term fluctuations, and short-term shocks. By enabling the decomposition of trend, seasonality, and noise components, this approach has become a powerful tool, not only for descriptive analyses, but also for forecasting models (Rhif et al., 2019). Wavelet modeling may be likened to how historians study the past across different temporal scales. A historian may examine long-term developments such as the rise and fall of empires over centuries, while also analyzing medium-term patterns like economic periods of reform within a few decades. At the same time, they may focus on short-term events—such as wars, revolutions, or political crises—that occur over months or years but can have lasting impacts. Wavelet analysis allows all these historical layers to be examined simultaneously. It distinguishes between enduring structural trends, recurring cyclical patterns, and sudden disruptive events, while also showing when each of these processes becomes influential. Similarly, a librarian may first look at annual statistics to understand overall trends in journal or database usage. Then, they may zoom into monthly or weekly data to identify seasonal patterns, such as increased usage during exam periods. Finally, they might examine daily spikes to detect sudden events, such as a surge in access following a major assignment. Wavelet modeling may be used to analyze library usage, and may facilitate the detection of regularities, cycles, and sudden shocks in data types, such as circulation data, user queries, and access frequencies, as well as the classification of digital content. Moreover, wavelet modeling may be used for decomposition of usage data before conducting machine learning–based econometric models, in order to effectively conduct forecasting models.
This study primarily conducts continuous wavelet coherence analyses, employing Morlet and Mexican Hat (Ricker) mother wavelets to identify time–frequency relationships between e-resource usage and exogenous variables that may influence such usage. Subsequently, the study develops a forecasting framework by conducting seven regression algorithms strengthened with Safe Bayesian Optimization and forward-chaining cross validation on usage data decomposed by Maximal Overlap Discrete Wavelet Transform (MODWT) and Multiresolution analysis (MRA). Finally, the study employs SHapley Additive exPlanations (SHAP) to identify the most influential cycles and factors affecting the forecasts. This framework may provide libraries the possibility to move beyond aggregate collection decisions to more differentiated management strategies. Traditional metrics depending on total downloads or cost-per-use are usually insufficient for determining the value of e-resources. The proposed approach enables a more detailed evaluation of how and when resources generate value. For example, identifying a resource with moderate usage but strong and recurring usage peaks during critical academic periods may be particularly valuable. In many academic libraries, database usage typically peaks during midterm and final examination periods, while certain databases experience short-lived surges following assignment announcements or thesis deadlines. A wavelet-based approach distinguishes these recurring academic cycles from one-off spikes, such as temporary increases caused by promotional access or course-specific requirements, thereby preventing misleading interpretations of demand. During the COVID-19 pandemic, many libraries observed abrupt increases in e-book and remote database usage; however, not all of these increases were translated into long-term demand. By identifying whether such changes persist at medium- or long-term scales, libraries can make more informed decisions about whether to maintain expanded electronic collections, or not. Furthermore, providing usage forecasts with multi-scale dynamics strengthens libraries’ bargaining positions and supports more sustainable licensing strategies. If a database shows declining long-term usage despite short-term fluctuations in demand, librarians may justify subscription reductions, or the renegotiation of pricing terms. Conversely, resources having strong and persistent medium- and long-term demand may be prioritized in negotiations. In the read-and-publish agreements, where costs are linked to both access and publishing activity, libraries may prioritize agreements with structurally strong portfolios, and keep others on read-only, or pay-as-you-publish models, improving cost-per-value.
This study contributes to the field of library and information science (LIS) by offering a robust, data-driven methodology that supports more flexible, efficient, and evidence-based collection management decisions.
Literature review
Early signal processing techniques
Signal processing and wavelet-based approaches in the field of LIS have been applied in a limited number of studies, and across different contexts. In early research, the presence of weekly and seasonal cycles in library usage was demonstrated through circulation statistics using spectral and Fourier-based methods (Decroos et al., 1997; McGrath, 1996). Subsequent studies have employed wavelet and other signal processing techniques primarily in the contexts of multimedia content analysis, digital collections, text classification, and information retrieval (Darányi et al., 2012; Rui et al., 1999; Shukla and Das, 2022; Wang et al., 2006). Research indicates that signal processing approaches provide applicable and effective tools for analyzing complex patterns in the field of LIS.
However, existing research has largely relied on the assumption of stationarity and has not adequately addressed the multiscale and time-varying nature of library usage. In particular, these approaches have been limited in capturing how usage patterns evolve in response to external influences. Similarly, applications of wavelet and related techniques in multimedia and digital collections have mainly focused on feature extraction and content representation, rather than on modeling temporal usage dynamics. In contrast, our research adopted a time–frequency and multiscale analytical framework to examine the non-stationary nature of library usage data by explicitly incorporating exogenous variables such as pandemic effects, economic conditions, and institutional cycles.
Classical time series models and regression algorithms
Predictive studies in the field of LIS are relatively limited (Chu, 2015; Togia and Malliari, 2017; Turcios et al., 2014; Zhang et al., 2017; Zhang and Tian, 2023). The most common approaches in this line of research are classical models based on time series analysis (Kotu and Deshpande, 2019, p. 395). In early studies, library usage data were generally forecast using methods such as trend analysis, exponential smoothing, and moving averages; most of these models produced strong forecasts (Ahmadi et al., 2013; Aliu and Nwankwo, 2019; Marasinghe, 2020). Autoregressive Integrated Moving Average (ARIMA) and seasonal ARIMA models were the most frequently used methods. These models were shown to predict both library usage statistics and bibliometric indicators with low error rates (Esh and Ghosh, 2023; Kumar and Alpha Raj, 2016; Song and Cao, 2022). Although such studies on library usage statistics demonstrated that library usage data contained predictable structures, they remained limited in addressing the multiscale and non-stationary nature of library usage data. The limitations of classical time series methods in the presence of abrupt external shocks such as the COVID-19 became more evident. In a study that forecast the use of health databases using a Centered Moving Average (CMA) model, based on usage data from January, 2016 to March, 2020, actual usage in April, 2020 was observed to be significantly higher than the predicted values; this deviation was suggested by the authors to have been associated with the increase in COVID-19–related research at the onset of the pandemic (Sharpe and Evans, 2022).
Regression analysis is one of the commonly used methods in forecasting studies examining library usage such as e-book features, database usage, and interlibrary loan requests through statistical models (Grabowsky et al., 2020; Kohn, 2018). In addition, machine learning–based models made significant contributions to forecasting processes in LIS (Daimari et al., 2023). Predictive frameworks based on methods such as deep neural networks (DNN), support vector machines/regression (SVM/SVR), and random forests (RF) were successfully applied to predicting book usage patterns (Iqbal et al., 2020). These predictive models were also used to support book acquisition decisions (Wu et al., 2022). In studies conducted within the context of demand-driven acquisition (DDA), methods such as random forests, AdaBoost, and logistic regression were shown to be effective in predicting the future use of information resources (Jiang et al., 2019; Walker, 2021; Walker and Jiang, 2019). Furthermore, hybrid approaches combining classical time series models and machine learning methods have also been explored. A study that combined ARIMA and support vector regression (SVR) reported that the hybrid model achieved higher accuracy than individual models (Pan et al., 2021). There have also been studies in which library usage has been examined together with other data sources, such as web data and social media, rather than solely through circulation statistics (Al Baghal, 2019).
The literature indicates that classical time series models and machine learning techniques provide valuable insights for forecasting library usage. However, collection decisions—such as subscriptions, licensing models, and budget allocations—may still lack the flexibility required to adapt effectively to evolving user demand. For example, a school library may purchase multiple copies of a e-resource following a short-term spike driven by a specific assignment, only to find that demand declines once the assignment ends, resulting in underutilized materials. Similarly, in a public library, event-driven increases in demand for social support or employment resources may not be adequately captured by models based on historically stable patterns. Wavelet-based signal processing techniques address these limitations by analyzing usage data across multiple time scales. By decomposing data into short-, medium-, and long-term components, they distinguish temporary fluctuations from sustained trends. This is specifically valuable in library settings shaped by academic cycles and external shocks, enabling more flexible collection management. Thus, libraries might develop more proactive, adaptive, and evidence-based collection strategies rather than reactive adjustments.
Methodology
This study examines and forecasts e-resource usage at the Library of the Grand National Assembly of Türkiye (GNAT) using usage statistics from OECD iLibrary, one of its most heavily used databases. Wavelet coherence analysis was employed to identify scale-dependent relationships between usage patterns and exogenous factors. Forecasting was conducted using wavelet transform–based regression models supported by Safe Bayesian Optimization.
Data collection and preparation
The OECD iLibrary, which was subscribed to by the GNAT library between 2016–2024, provided full-text resources across key policy and economic domains and was intensively used during the Government of Türkiye's budget preparation period. Due to limited historical depth in COUNTER Release 5 reports, analyses were conducted using COUNTER Release 4 statistics, specifically Journal Report 1 (JR1) and Book Report 1 (BR1). Monthly usage data were collected, and forecasting analyses were performed in the R environment.
Given that library usage reflects temporal and contextual influences, the analyses incorporated selected exogenous variables alongside journal and book usage data to better capture external drivers of observed trends.
Construction of fuzzy dummy variables
In order to capture the temporal influence of major structural and institutional events on e-resource usage, four fuzzy dummy variables were constructed to represent the COVID-19 pandemic, the economic crisis, legislative periods, and budgeting cycles. Rather than binary indicators, these variables were modeled as continuous values between 0 and 1, reflecting the gradual onset, persistence, and decline of each event's influence over time.
The COVID variable models the institutional and behavioral impact of the pandemic between early 2020 and mid-2022. A membership function gradually increased during the outbreak's escalation (March–June 2020), remained at peak intensity during strict lockdowns (mid-2020 to mid-2021), and then progressively declined as in-person operations resumed.
The Crisis variable represents the macroeconomic instability beginning in 2018 and its prolonged effects. The function increases during the escalation of financial stress in mid-2018, peaks during heightened volatility in late 2018–mid-2019, and then gradually declines while accounting for continued exchange rate pressures through subsequent years.
The Legislative variable models the cyclical influence of parliamentary sessions (October–July). Membership values increase at the start of each legislative year, peak during intensive policy and debate periods (November–June), and approach zero during recess (July–September), reflecting continuous fluctuations in policy-related information demand.
The Budget variable captures the national budgeting cycle, modeled using a triangular fuzzy function. Membership values rise from late summer, peak during October–December budget negotiations and approvals, and decline as the fiscal year begins, reflecting the temporal concentration of budget-driven information needs.
Sensitivity analysis, a tool for assessing the impact of changes in variables (Kahraman and Haktanır, 2024: 145), was applied to measure the effectiveness of fuzzy dummy variables on the model outcomes. Six scenarios were implemented by systematically redefining each fuzzy dummy variable to capture assumptions about intensity, persistence, and timing of external effects:
Baseline specification reflecting the original fuzzy construction; Fast-decay transformation emphasizing short-lived, rapidly diminishing effects; Slow-decay transformation capturing more persistent and gradual influences; Forward temporal shift (+2 periods) to model delayed impacts; Backward temporal shift (−2 periods) to account for anticipatory or leading effects; Binary (crisp) transformation which removes gradation, and represents a traditional dummy-variable approach.
Morlet wavelet coherence results were highly robust for journal and book usage, particularly in the medium term. Short-term coherence remained stable except for economic crisis sensitivity. Mexican hat coherence showed strongest long-term stability for legislative and budgeting variables. ElasticNet, Lasso, and Ridge remained consistently reliable, whereas SVR showed instability under alternative fuzzy specifications.
Wavelet transform
Wavelet transform is a signal processing technique that enables the analysis of time series simultaneously in the time and frequency domains, and is particularly suitable for non-stationary data (Hadi and Tombul, 2018). By decomposing a series into multiple scale-dependent components, wavelet analysis allows the identification of both short-term fluctuations and long-term patterns, while preserving information about when these dynamics occur (Akujuobi, 2022: 46). This time–frequency localization makes wavelet-based methods especially effective for examining dynamic signals characterized by structural changes and transient behaviors (Kayral et al., 2025). For a time series x(t), the mother wavelet ψ(t) serves as the generating function for a family of wavelets obtained by scaling (s) and translating ψ(t):
Descriptive analyses with continuous wavelet transform
Within the continuous wavelet transform (CWT) framework, wavelet coherence analysis extends standard time–frequency analysis by examining the localized correlation between two time series across both time and frequency domains. Rather than analyzing each signal separately, wavelet coherence identifies when and at which periodicities two variables co-move, and whether one leads or lags the other. The Morlet wavelet coherence is commonly used to detect oscillatory and periodic relationships, providing high resolution for identifying co-movements and phase relationships among time series. By contrast, the Mexican Hat wavelet coherence, derived from the second derivative of a Gaussian function, is more sensitive to structural changes, peaks, and long-term features, making it particularly useful for highlighting persistent or non-oscillatory coherence patterns, and emphasizing broader, smoother relationships. Together, within the CWT framework, Morlet coherence captures time-varying cyclical interactions, while Mexican Hat coherence emphasizes structural and long-term alignment, enabling a comprehensive interpretation of both transient and enduring relationships between variables (Kumar and Kumar, 2020).
Before applying wavelet coherence, wavelet power analysis was conducted to examine the energy density of the JR1 and BR1 series over time and to identify dominant periodic components. The resulting wavelet power scalograms highlight significant periodicities and their temporal evolution (Parmar and Chothodi, 2025). Wavelet Coherence (WTC) analysis was then applied to examine the dynamic relationships between JR1 and BR1 and exogenous variables across both time and frequency domains. Morlet-based wavelet power and coherence analyses were implemented using the WaveletComp package. The Morlet wavelet was preferred due to its strong time–frequency localization, enabling the detection of recurring academic usage patterns. The selected 2–36-month period range corresponded to library-relevant cycles, including short-term fluctuations (e.g., assignment deadlines or committee meetings), medium-term dynamics (e.g., academic or legislative cycles), and long-term structural trends (e.g., subscription or policy effects). The analysis employed parameters of dt = 1 (monthly frequency), dj = 1/12 (high spectral resolution), and loess.span = 0.3 (balanced smoothing), ensuring sensitivity to local variations while preserving interpretability. Statistical significance was assessed using Monte Carlo simulations (n.sim = 2000) with 2000 permutations. Wavelet coherence scalograms were interpreted by focusing on statistically significant regions within the cone of influence (COI), where edge effects are minimized. High coherence regions, represented by warm colors, indicate strong co-movement, while cooler colors denote weak relationships. The vertical axis reflects time scales, and the horizontal axis represents temporal evolution, enabling simultaneous analysis of multi-scale dynamics (Roesch and Schmidbauer, 2025). Phase arrows indicate the direction and timing of relationships: rightward arrows denote positive synchronization, leftward arrows indicate negative relationships, while diagonal arrows capture lead–lag dynamics between the variables (Vacha and Barunik, 2012). In addition, Mexican Hat wavelet power and coherence analyses were implemented using the PyWavelets library in Python via the reticulate package in R. Complementing the Morlet approach, the Mexican Hat coherence analysis captures interactions, such as crisis-driven usage changes, making this approach particularly suitable for identifying the impact of localized and abrupt changes (Nason, 2008: 79).
To compare the results of the Morlet and Mexican Hat wavelet coherence analyses, the wavelet correlation (band-average coherence) was conducted using the W2CWM2C package by averaging coherence values within 2–4, 4–8, 8–16, 16–32, and 32–36-month ranges, restricted to statistically significant regions within the COI.
Predictive analyses with discrete wavelet transform
For predictive analyses, the 108-month datasets (JR1 and BR1) were divided chronologically into a training period (2016–2022) and a testing period (2023–2024). The training period was used for model construction and hyperparameter optimization, and for evaluation through cross-validation, while the testing period was used to assess out-of-sample predictive performance. The analyses were conducted following decomposition through Maximum Overlap Discrete Wavelet Transform (MODWT) based Multiresolution Analysis (MRA), which decomposes time series into multi-scale components without information loss, while preserving temporal alignment and shift invariance. MODWT is a time-series decomposition method that breaks a signal into components associated with different time scales such as short-, medium-, and long-term fluctuations without downsampling the data. Unlike the standard discrete wavelet transform, MODWT preserves the original time alignment, making it especially suitable for irregular, non-stationary series such as library usage data. MRA builds on MODWT by reconstructing the original series into additive components—detail levels capturing the variations at specific scales and a smooth component representing the long-term trend—so that each time point can be interpreted in terms of its scale-specific contributions. When used together, this approach first decomposes the data into scale-based components and then recombines it into interpretable layers. This enables researchers to link observed behavior to distinct temporal dynamics and supporting both explanation and forecasting (Habimana, 2017). In this study, for each series, MODWT + MRA were applied using the least asymmetric Daubechies 8 wavelet due to its balanced time–frequency localization and suitability for nonlinear economic time series. Five decomposition levels were applied to reduce non-stationarity and improve forecasting stability, capturing oscillations up to 32 months and separating short-term (D1–D2), medium-term (D3–D4), long-term (D5), and scaling (S5) components. These wavelet-derived features were then combined with the exogenous variables (covid, legislative, budget, crisis). Based on forecasting trials using lagged usage data (e.g., 1–6, 1–8, 1–12, and 1–24 months), lags of 1–6 months were identified as the most consistent predictors. Accordingly, 1–6 month lags were constructed as predictors for all wavelet components, as well as for exogenous variables (COVID, legislative, budgetary, and crisis-related factors). Seasonality was modeled using a 12-month frequency, and all decompositions preserved the original time-series length, ensuring complete reconstruction and a stable multiscale representation.
Seven regression models were used to capture different patterns in library usage data. First, Support Vector Regression (SVR) with a radial basis function (RBF) kernel was applied to model complex and nonlinear relationships, such as seasonal usage patterns, crisis-related changes, and policy effects. To provide comparison, three linear models—Ridge, Lasso, and ElasticNet—were implemented using the glmnet package. These models help address issues such as multicollinearity, and large numbers of predictors. Ridge focuses on stability, Lasso performs automatic variable selection, and ElasticNet combines both approaches to handle correlated variables effectively. Gaussian Process Regression (GPR) was also used to model the data in a flexible way. This model is particularly useful for capturing both gradual changes and sudden shifts in usage patterns. In addition, k-Nearest Neighbors (kNN) regression was included as a nonparametric baseline model, capturing local similarities in the data. Decision Tree regression was also applied to identify nonlinear and rule-based relationships, such as threshold effects in usage behavior.
Regression models were tuned using a two-stage procedure. Safe Bayesian Optimization was employed to explore bounded hyperparameter spaces using an Upper Confidence Bound acquisition function (κ = 2.576). When the optimization process failed, a caret-based fallback using forward-chaining cross-validation was applied to preserve temporal ordering and prevent look-ahead bias. All validation folds were constructed sequentially to maintain temporal integrity, with a minimum of 24 training observations per fold to ensure estimation stability given the lagged feature structure. All input data were standardized to ensure consistency across variables. Model tuning procedures were designed to reduce overfitting and improve generalization performance, while, in the case of Decision Tree regression, additional constraints were applied to ensure meaningful data partitions.
Model evaluation
Predictive performance of models was evaluated using multiple accuracy metrics, including the coefficient of determination (R2), root mean squared error (RMSE), mean absolute error (MAE), symmetric mean absolute percentage error (SMAPE), and mean absolute scaled error (MASE). Together, these measures provide a multidimensional assessment of forecast accuracy.
In addition to accuracy assessment, model interpretability was enhanced through SHAP (SHapley Additive exPlanations) analysis. SHAP is a method that explains how machine learning models generate predictions by quantifying the contribution of each variable, based on principles from Game Theory (Alomari and Andó, 2024). In this framework, a prediction is considered as the combined outcome of multiple “players,” such as usage levels, economic conditions, and institutional variables, each contributing to the final result. This approach quantifies the marginal influence of predictors—including wavelet-decomposed components, lagged features, and exogenous variables—on forecast outcomes. Integrating SHAP analysis with traditional error metrics allows not only the identification of high-performing models but also a transparent understanding of the drivers underlying predictive performance, thereby strengthening the interpretability and reliability of the forecasting framework.
Results
Wavelet power and Mexican hat wavelet power
The results of the wavelet power and Mexican hat wavelet power analyses conducted on the OECD iLibrary journal report (JR1) and book report (BR1), consisting of monthly observations between 01.01.2016–31.12.2024 are shown in Figure 1.
Wavelet power and Mexican hat wavelet power: Journal and book usage. Note: A Morlet wavelet power scalogram examines the distribution of energy (variance) of a single time series across time and frequency: the horizontal axis represents time, the vertical axis shows the period (scale), and the color intensity indicates power, with hot colors reflecting stronger oscillatory activity at a given time and scale. It is particularly suitable for identifying stable and evolving cyclical patterns, with short-, medium-, and long-term periodicities appearing at different bands; statistically significant regions are often enclosed by contours, while the cone of influence marks areas affected by edge effects. A 3D Mexican Hat power surface represents power through surface height and color, where peaks indicate localized concentrations of energy. This highlights bursts, turning points, and structural shifts, rather than smooth oscillations. Broad elevated regions suggest persistent structural variation, while sharp peaks indicate short-lived, event-driven dynamics (Bolós and Benítez, 2022).
The journal usage wavelet power scalogram indicated that journal usage was primarily led by short- and medium-term dynamics rather than long-term structural cycles. The strong and persistent power in the 8–16-month band suggested the presence of recurring medium-term usage patterns, probably associated with institutional rhythms such as academic calendars and research cycles. At the same time, the fragmented activity in the 2–6-month band reflects short-term volatility, driven by episodic events, changing user behavior, or temporary access conditions. The log10-transformed Mexican Hat power surface of journal usage further suggested that journal usage could experience temporary surges that built upon existing medium-term cycles in late 2023 and early 2024. The absence of significant power at longer periods (>16–20 months) indicated that journal usage did not follow stable long-term cyclical behavior, but instead evolved through a combination of recurring medium-term structures and irregular short-term fluctuations.
The book usage wavelet power scalogram indicated that book usage was governed by a combination of stable medium-term cycles and pronounced multi-year structural dynamics, with relatively limited influence from short-term fluctuations. The persistent power in the 8–16-month band suggested regular, recurring patterns of use, possibly linked to institutional cycles. The temporary intensification in the 3–6-month and 2–4-month bands—especially in the years 2019–2021, and again in the years 2023–2024—reflected episodic increases in book usage, possibly associated with disruptions such as the pandemic, shifts to remote access, and short-term changes in access conditions. The strong and consistent power in the 25–36-month range in the Mexican Hat surface indicated that book usage was deeply influenced by long-term, multi-year cycles, suggesting a structural and cumulative pattern of engagement. These longer cycles likely reflected broader processes such as program development, research agendas, and institutional investment in learning resources. The stability of these multi-year oscillations through 2020–2023 implied that, unlike journal usage, book usage exhibited stronger long-term continuity and persistence.
Wavelet coherence and Mexican hat wavelet coherence
In this part of the research, time-frequency analyses across multiple temporal scales were conducted, in order to examine how journal and book usage responded to the key exogenous drivers: COVID-19, the economic crisis, the legislative period, and the budgeting period. Standard Morlet-based wavelet coherence was first applied, in order to identify scale-dependent synchronization, lead–lag structures, and alternating phases of positive and negative co-movement over short-, medium-, and long-term bands (Figure 2). Complementary Mexican-hat wavelet coherence surfaces were then estimated, in order to highlight smoother, more persistent medium- and long-horizon relationships, which were not visible in the Morlet results, thereby revealing deeper structural linkages between usage patterns and institutional or macroeconomic cycles (Figure 3). Finally, band-average wavelet coherence (wavelet correlation) measures were computed for both wavelet analyses, in order to summarize the strength of coherence within each key frequency band, and to allow a direct comparison of short, medium-, and long-term coherences across all usage-related pairs.
Wavelet coherence: Journal usage and exogenous variables. Note: In the wavelet coherence scalogram, the horizontal axis represents time, while the vertical axis indicates the period (frequency/time scale). In a 36-month scalogram, short-term cycles occur at 2–8 months, medium-term cycles at 8–16 months, and long-term cycles at 16 + months. The color scale (0–1; blue to red) reflects the strength of coherence, with hot colors indicating stronger relationships and cooler colors indicating weaker ones; statistically significant regions are often enclosed by a white contour. Phase arrows show the temporal structure of the relationship: rightward indicates positive co-movement, leftward negative co-movement; upward and downward indicate a 90° phase difference where the first variable leads the second and vice versa; northeast and southeast indicate in-phase movement with the first and second variable leading, respectively; northwest and southwest indicate anti-phase movement with the second and first variable leading, respectively (Vacha and Barunik, 2012). A 3D Mexican Hat wavelet coherence surface is read by combining height and color: the horizontal axis shows time, the depth/vertical axis shows period (scale), and the surface height indicates the strength of coherence. Elevated plateaus and peaks with warm colors represent strong co-movement, while valleys with cooler colors indicate weak relationships (Kirgo et al., 2021).
Wavelet coherence: Book usage and exogenous variables. Note: Interpretation of the wavelet coherence scalogram and the Mexican Hat wavelet coherence surface follows the conventions described in Figure 2 .
Journal usage and exogenous variables
There was a broad and statistically significant high-coherence between journal usage and the COVID-19 pandemic from late 2019 to approximately mid-2022, according to the wavelet coherence analysis. In Figure 2, the strongest coherence was observed in the 8–16-month band, with additional short-term peaks in the 4–6-month band between mid-2019 and mid-2021. Although a weak inverse relationship between the two variables was observed in the 6–8-month band during late 2019–2020, journal usage generally exhibited a direct relationship with pandemic dynamics from late 2019 to 2021. Similarly, around the 10–12-month band a direct and gradual co-movement was observed between the two variables until late 2021. From early 2022 to mid-2023, journal usage generally changed, depending on the effects of pandemic. The Mexican Hat wavelet coherence surfaces indicate that pronounced high-coherence regions between journal usage and the pandemic emerged primarily after 2021, with the strongest synchronization occurring in a medium to long term period. Coherence was the highest within the 30–36-month band, especially between September, 2022, and January, 2023, reflecting persistent long-term alignment of the two variables. These results suggested that the relationship between journal usage and the pandemic evolved into a stable long-term coherence over time, indicating that the pandemic sustained structural effects on journal usage beyond the initial shock period.
While wavelet coherence analysis revealed weak and fragmented coherence between journal usage and the economic crisis during 2018–2019, 2022, and 2024 in the short term—possibly driven by sudden economic fluctuations—a strong and persistent coherence region at the 8–16-month periodicity from late 2017 to 2021 suggests that longer-term macroeconomic instability influenced journal usage more strongly than short-term shocks. Until mid-2020, journal usage and the economic crisis exhibited an inverse relationship. After 2021, medium-term coherence weakened markedly, and irregular phase patterns during 2022–2024 indicated that the crisis was no longer coherent with journal usage. The weakening of coherence after 2021 implies that the relationship between the two variables became more unstable and less systematic in later periods. The Mexican Hat wavelet coherence surface revealed a pronounced high-coherence ridge at the 30–36-month periodicity during mid-2022, late 2023, and mid-2024, while distinct low-coherence regions appeared at the 2–6-month periodicity, particularly in late 2023 and early 2024, indicating strong long-term synchronization between journal usage and the crisis under intensified macroeconomic instability in Türkiye. Generally, the presence of strong long-term and weak short-term coherence confirmed that journal usage responded more to persistent macroeconomic stress than to transient economic fluctuations.
The wavelet coherence analysis between journal usage and the legislative period revealed the highest coherence region from 2017 to 2022, concentrated in the 8–16-month band, indicating a stable medium-term synchronization with changes occurring in journal usage–probably because of preparatory research before legislative sessions–before the beginning of legislative activity. In addition, occurrence of strong and sustainable medium-term coherence of the two variables especially between mid-2020 and mid-2021 might suggest the increasing demand for e-resources, because of pandemic-driven effects. The weakening of coherence between mid-2023 and early 2024 (during the parliamentary election preparation period in Türkiye, when parliamentary sessions were interrupted) indicated that journal usage was directly proportional to legislative work. According to the results of the Mexican hat wavelet coherence analysis, co-movement between journal usage and the legislative period were primarily strong at long-term scales. The strongest synchronization between the two variables occurred between mid-2022 and 2024 in the 28–36-month band, notably around July, 2022, November, 2022, July, 2024, and August, 2024. Continuous co-movement and high coherence between journal usage and a legislative period, which occurred particularly after 2022, suggested that usage patterns were permanently driven by institutional legislative dynamics.
Wavelet coherence analysis between journal usage and the budgeting period identified a dominant and persistent high-coherence region from 2017 to mid-2021 at the 8–16-month periodicity, indicating stable medium-term synchronization between the two variables. The positive co-movement between journal usage and the budgeting period during 2017–2021—suggesting that usage demand was shaped by recurring parliamentary budgeting processes—reversed during 2022–2024. This reversal points to a transformation in the underlying dynamics, whereby external pressures—such as post-pandemic adjustments, economic constraints, or changes in access and subscription strategies—may have disrupted the previously synchronized pattern, leading to a reconfiguration of how budgeting processes influenced journal usage in the short and medium term. The Mexican hat wavelet coherence analysis between journal usage and the budgeting period revealed a strong medium- to long-term synchronization across 2016–2024. The most pronounced coherence ridge emerged in 2023–2024 within the 30–36-month band. The results indicated that journal usage was systematically coherent with the longer-term legislative cycles, implying that decisions and negotiations within the parliamentary budgeting process permanently shaped journal usage behavior in the long-term.
Book usage and exogenous variables
The wavelet coherence analysis between book usage and the COVID-19 pandemic (Figure 3) revealed significant high coherence in the 8–16-month band from late 2019 to mid-2021, indicating a stable medium-term synchronization between the two variables. A secondary coherence region in the 3–8-month band during 2019–2021 captured short-term, event-driven fluctuations associated with lockdowns, parliamentary interruptions, and reopening phases, suggesting that the pandemic generated both episodic and cyclical effects on book usage. During 2020–2021, book usage exhibited an inverse relationship with pandemic dynamics, however, by this relationship became more direct, suggesting a delayed response shaped by changing access conditions, evolving user needs, and increased reliance on digital formats. The Mexican hat wavelet coherence further revealed strong long-term coherence in the 24–36-month band from shortly after the onset of the pandemic through 2021–2024, peaking between mid-2021 and late 2023. This persistent coherence indicated that book usage became aligned with the structural phases of the pandemic, reflecting sustained e-resource usage and hybrid work regimes.
A dominant high-coherence region between 2018 and 2022 within the 8–16-month band was observed by the wavelet coherence analysis between book usage and the economic crisis, with a short-term coherent cluster in 2024 within the 2–6-month band. These results identified that book usage was more strongly consistent with the effects of the economic crisis over the medium-term. In the 2018–late 2020 period, changes in the crisis dynamics preceded and drove adjustments in book usage, pointing to a leading role for macroeconomic conditions in book usage and reflecting periods of both inversely and directly proportional responses between book usage and the crisis. However, during late 2021–2022, book usage began to respond more systematically and in alignment with crisis-related effects. This period might have reflected the simultaneous effects of the pandemic and the crisis on usage. In 2024, within the 2–6-month range, this relationship varied frequently. The Mexican hat wavelet coherence surface showed a strong and persistent coherence between book usage and the crisis across the 2016–2024 period. The highest-coherence regions appeared mainly within 12–20-month cycles in late 2021–2022, reflecting strong synchronization between the two variables during phases of intensified crisis dynamics. Coherence, moreover, remained high through 2023–2024 at 12–15-month periods. These results indicated that crisis led to long-lasting, stable changes in book usage both in medium-and long-term periods.
The wavelet coherence between book usage and the legislative period was highest within the 8–16-month band from mid-2016 to 2024, indicating a strong and stable medium-term relationship between book usage patterns and legislative activity. A secondary high-coherence region in the 4–8-month band during 2019–2021 reflected synchronized short-term oscillations. The two variables moved in direct proportionality in the 8–16-month band, in which changes in book usage occurred before the periods of intensified parliamentary activity. This reflected the more intense usage of e-books in preparation for legislative sessions. The Mexican hat wavelet coherence analysis, on the other hand, revealed a high and persistent long-term synchronization within the 20–36-month band, indicating a structural multi-year alignment between legislative activity and book usage, reflecting sustained institutional information-seeking, linked to law-making cycles, parliamentary sessions, and preparatory phases. Long-term coherence continued through 2022–2024, demonstrating near-perfect alignment between legislative activity and book usage.
Likewise, the highest wavelet coherence between book usage and the budgeting period was observed at the 8–16-month periodicity from mid-2016 to 2024. During this period, the budgeting cycle led to book usage, with a positive co-movement between the two variables. This situation pointed to a strong synchronization between book usage and budget preparation, review, and approval cycles in the medium term. The Mexican hat wavelet coherence surface identified peak coherence values in the 24–36-month band, reflecting strong and persistent coherence between book usage and multi-year budgeting and approval cycles. A large, contiguous high-coherence region revealed a systematic cyclical relationship between the two variables, with coherence intensifying around budget preparation phases in late 2021, mid-2022, mid-2023, and mid-2024, coinciding with parliamentary discussions. These results implied that book usage became systematically compatible with institutional financial planning processes, particularly during periods of intensified parliamentary debates.
Findings of wavelet coherence between journal and book usage and exogenous variables, suggested that in the short term (2–6 months), coherence of e-resource usage with the pandemic and economic crisis was fragmented and episodic, with frequently shifting lead–lag directions, reflecting abrupt lockdowns, reopening phases, sudden currency shocks, and rapid policy responses. Medium-term dynamics (8–16 months) showed the most stable synchronization, indicating cyclical adjustment processes driven by sustained pandemic conditions, prolonged macroeconomic instability, regular legislative calendars, and annual budget preparation cycles. In contrast, long-term coherence (16–36 months) revealed strong and persistent synchronization, reflecting structural transformations, such as hybrid work regimes, cumulative economic stress, multi-year legislative agendas, and institutional budgeting rhythms.
In order to provide a comprehensive assessment, band-averaged wavelet coherence (wavelet correlation) results are presented below, derived from wavelet coherence and Mexican hat wavelet coherence analyses. The standard wavelet coherence band-average results emphasized that journal and book usage exhibited their strongest coherence with the legislative and budgeting periods within the 6–12-month range, indicating short- and medium-term institutional cycles shaping usage behavior. Book usage demonstrated high coherence values of 0.86–0.87, while journal usage showed slightly weaker, but still substantial coherence of approximately 0.69. In contrast, the Mexican hat wavelet coherence band-average analysis revealed notable long-term synchronization in the 12–36-month band, with coherence values ranging from 0.74 to 0.99, and reaching 0.99 in the 24–36-month band for both journal and book usage. Within the 6–12-month range, book usage exhibited particularly strong coherence with the legislative and budgeting periods, with values of 0.94–0.95. In the 2–8-month bands coherence changed between 0.50 and 0.80, whereas around the 16-month band it ranged between 0.74 and 0.83. The coherence values in the standard wavelet analysis usually varied between 0.40 and 0.65, suggesting that short-term increases in e-resource usage were temporary and unstable. The Mexican hat wavelet analysis results displayed that the exogenous variables affected journal and book usage with relatively high and persistent coherence values (Figure 4).
Comparison of wavelet coherence and Mexican hat wavelet coherence results.
Predictive analyses
The MODWT + MRA decomposition results of journal usage, obtained using the la8 wavelet filter and a decomposition depth of J = 5 (hyperparameter controlling the number of scales), revealed that journal usage variability was predominantly driven by high-frequency components, with D1 and D2 explaining 58.6% and 11.5% of the total variance. These components capture very short-term fluctuations, indicating that journal usage is primarily influenced by immediate and rapidly changing dynamics. Medium-scale components (D3 and D4) accounted for 13.9% and 7.6% of the variance, representing intermediate periodicities, while the low-frequency components (D5 and the smooth component S5), which reflected long-term structural trends, contributed relatively little (6.3% and 2.2%). The reconstruction plot showed close alignment between the original journal usage series and the reconstructed signal from D1–D5 and S5, confirming that the la8 wavelet filter and J = 5 decomposition effectively preserved the essential temporal structure of the data (Figure 5).
MODWT + MRA components with actual and reconstructed series of journal usage.
A similar pattern was observed for book usage, where D1 (44.7%) and D2 (20.0%) dominated the variance, followed by moderate contributions from D3 (24.4%) and D4 (4.1%), and limited influence from long-scale components (D5 and S5). The reconstruction comparison between the original and reconstructed book usage series showed a close alignment, indicating that the MODWT + MRA decomposition accurately preserved the essential temporal structure of the data, and effectively captured the timing and magnitude of major peaks, troughs, and fluctuations. This consistency confirmed that the chosen wavelet filter and decomposition depth were methodologically sufficient to preserve the essential temporal structure of the book usage data (Figure 6).
MODWT + MRA components with actual and reconstructed series of book usage.
For journal usage forecasts, the comparison between actual and forecast values demonstrated that the ElasticNet, SVR, Lasso and Ridge models generally followed the overall trend of the actual journal usage values. Decision tree, GPR, and kNN demonstrated weaker adherence to the actual data. Decision tree and kNN showed visible underestimation during months of sharp increases, and produced smoother forecast trajectories, which failed to fully reflect abrupt high-frequency shifts in the journal usage values. Journal usage was characterized by high-frequency fluctuations, which could only be accurately captured by models capable of preserving short-term variability, while maintaining overall trend structure. The strong performance of ElasticNet, Ridge, Lasso, and SVR suggested that these models effectively balanced flexibility and regularization, allowing them to track both gradual trends and sudden increases in usage. In contrast, the underperformance of Decision Tree and kNN models, which produced oversmoothed forecasts and underestimated peak values, indicated that methods lacking sufficient sensitivity to rapid changes were not well-suited for modeling library usage data. (Figure 7).
Forecasting results for journal usage.
Book usage forecasts indicated that Ridge, Lasso, ElasticNet, and SVR substantially outperformed the other models for book usage data. Ridge regression emerged as the best-performing model, while Lasso regression provided the second most accurate forecasts. By contrast, GPR, kNN, and decision tree regression models failed to produce consistent results. Similar to journal usage, book usage followed relatively stable and structured patterns, which were best captured by regularized linear models. The superior performance of Ridge and Lasso regression implied that the underlying relationships in book usage data were largely linear with moderate correlations, where shrinkage-based regularization effectively stabilized the forecasts. In contrast, the inconsistent performance of GPR, kNN, and decision tree models suggested that more flexible or nonlinear approaches were not suitable for modeling book usage patterns. (Figure 8).
Forecasting results for book usage.
For journal usage forecasts, ElasticNet demonstrated the strongest results, achieving both the highest R2 value (0.99) and the lowest RMSE (0.93), MAE (0.69), SMAPE (0.35), and MASE (0.10) among all models. These metrics indicate that ElasticNet produced forecasts closest to the observed data, while minimizing both variance and prediction errors. In addition, the SVR, ridge, and lasso models demonstrated comparatively strong performance, with R2 values of 0.89, 0.87, and 0.87, respectively. The GPR, kNN, and decision tree models produced very low R2 values, and so lacked model reliability. In the book usage forecasts, Ridge, Lasso, ElasticNet, and SVR produced more accurate forecasts, which closely tracked the distinct peaks and declines observed in the actual data, with R2 values 0.99, 0.98, 0.97 and 0.95 respectively. The Ridge regression provided the most accurate results with the lowest RMSE (1.34), MAE (1.02), and MASE (0.08) in comparison to the other models. The lowest SMAPE value (34.61) was obtained by the Lasso regression, which was the second most accurate forecast model for book usage data. GPR, kNN, and decision tree models did not provide consistent forecast results (Table 1).
Forecast accuracy and error comparison for journal (JR1) and book usage (BR1).
The strong performance of regularized linear models suggested that the underlying relationships in library usage data were mostly linear and stable. Ridge, Lasso, and ElasticNet relied on shrinkage, which appeared to be well-suited for the moderate correlations and structured temporal patterns of monthly usage statistics. This finding indicated that library usage followed largely predictable patterns through regular, smoothed, and consistent linear dependencies. Furthermore, the success of SVR, a nonlinear model, revealed that e-resource usage had some non-linear components.
In addition to error evaluations, SHAP values were computed using the fastshap package, in order to provide model-based interpretability, enabling feature importance rankings for each algorithm. The SHAP feature importance results indicated that journal usage forecasts were predominantly driven by short-term wavelet components. Through ElasticNet, Lasso, Ridge, and SVR, the very short-term components at lag 1 and lag 2 produced the highest SHAP values, confirming that high-frequency dynamics were the primary contributors to predictive performance. Oscillations at short- and medium-term scales, represented by short-term components at lags 1, 2, and 3 and the medium-term component at lag 1, also contributed substantially to the forecasts. In the Lasso model, very short-term component at lag 1 accounted for the largest contribution (about 25%). In the SVR forecasts, the economic crisis variable at lag 5 and lag 6 showed the strongest effects, while the Ridge model reflected a considerable contribution from the economic crisis variable at lag 6. The legislative period appeared only marginally in the Ridge model, exerting only minimal influence (Figure 9).
SHAP feature importance for journal usage forecasts.
Similar to journal usage forecasts, book usage forecasts were largely dominated by very short-term, high-frequency wavelet components. The very short-term component at lag 1 produced the highest SHAP contributions across the ElasticNet, Lasso, Ridge, and SVR models. Additional contributions came from short term components at lags 1 and 3, and from the medium-term component at lag 1, which influenced the forecasts at moderate to high levels. In the SVR model, the economic crisis variable at lag 6 had a strong impact, while contributing at a moderate-low level at lag 4. Ridge regression forecasts also reflected moderate-low effects of the crisis at lags 3 and 5 (Figure 10).
SHAP feature importance for book usage forecasts.
SHAP results showed that both journal and book usage forecasts were mainly driven by very recent usage behavior, with short-term components providing the strongest predictive power, while external factors had more delayed and secondary effects. Although short- and medium-term components contributed to improving forecast stability, their influence was weaker than the immediate past usage. The impact of the economic crisis in lags 4–6 indicated that macroeconomic conditions affected usage gradually. However, the relatively weak role of the legislative period suggested that institutional cycles shaped overall usage structure, but did not strongly influence short-term predictions.
Discussion
The wavelet-based signal processing analyses revealed that OECD iLibrary journal and book usage at the GNAT Library exhibited a pronounced multi-scale structure shaped by both institutional cycles and macroeconomic pressures. The coexistence of short-, medium-, and long-term frequency dynamics suggests that e-resource usage cannot be adequately explained by single-scale or purely event-driven interpretations.
The coherence analyses indicated that the COVID-19 pandemic generated fluctuating short- and medium-term relationships between 2019 and 2023. The initial short-term fluctuations in usage—such as rapid increases in digital access during pandemic lockdowns —reflected immediate behavioral responses. However, the emergence of strong long-term coherence between e-resource usage and the pandemic after 2021 suggested that these temporary adjustments evolved into permanent usage patterns, such as sustained reliance on e-journals, remote access, and hybrid service models. Many libraries expanded e-resource subscriptions during the pandemic as a temporary measure, yet continued these investments afterward according to the changes in user preferences. A similar interpretation may be applied to the economic crisis results. Users did not react instantly to the economic crisis, but instead gradually modified their behavior over time. This may reflect a shift whereby users reduced their reliance on paid or specialized resources and increased their use of essential or more widely accessible materials. Institutional cycles, such as legislative and budgeting periods, further reinforced these findings. E-resource usage, more importantly, was closely tied to institutional workflows and decision-making processes. Increased journal and book usage before legislative/budgeting periods reflected preparatory research activities. These results demonstrated that e-resource usage was not merely reactive, but dynamically embedded within institutional temporal structures.
The forecasting results complemented this interpretation by showing that, despite short-term volatility, usage patterns remained predictable and structured. The dominance of high-frequency components confirmed the importance of recent usage behavior, while the success of regularized models indicated that underlying relationships were stable and moderately correlated. This means that libraries may anticipate demand using short-term usage data, requiring real-time responsiveness to usage changes, such as changing access models when demand increases. The findings aligned with prior wavelet-based research in information retrieval and digital collections, which showed that multiresolution approaches outperform traditional models in capturing long-term structural rather than purely short-term patterns (Darányi et al., 2012; Rui et al., 1999; Shukla and Das, 2022; Wang et al., 2006). Studies on topic evolution and scholarly communication (Lee and Song, 2020; Zhu et al., 2018) similarly demonstrated cyclical and multi-scale dynamics. However, while earlier research primarily focused on textual or thematic evolution, this study extended the literature by empirically demonstrating that e-resource usage itself reflected layered temporal dependencies shaped by both institutional and macroeconomic processes. Forecasting analyses further supported the structural interpretation. The dominance of high-frequency components in variance decomposition indicated that short-term fluctuations drove immediate variability, yet medium- and long-scale components provided underlying stability. This leads to short-term surges in e-resource usage during parliamentary debates, which can be managed through rapid access adjustments, while long-term collection planning should be aligned with legislative cycles and institutional priorities. The superior performance of shrinkage-based regression models suggests that usage patterns are structured and moderately correlated, with limited nonlinear features. This supports the development of transparent, robust, and operational forecasting tools for routine collection management.
SHAP-based interpretability confirmed that forecasts were primarily driven by short-term wavelet components, while exogenous crisis indicators exerted selective longer-lag effects. These results collectively demonstrate that a multiscale modeling strategy offers both explanatory and predictive advantages over aggregate time-series approaches. Finally, the 108-month dataset enabled differentiation between transient disturbances and structurally embedded behaviors. The temporal depth of the dataset allowed the decomposition framework to capture layered dependencies that would remain obscured in shorter or aggregated datasets.
Conclusion
This study demonstrated that OECD iLibrary usage at the GNAT Library exhibited structured, multi-scale temporal dynamics shaped by external shocks and institutional cycles. By applying wavelet coherence and multiresolution forecasting techniques, the research revealed that e-resource usage reflected both immediate event-driven fluctuations and enduring structural alignments.
The results showed that pandemic-related and macroeconomic effects generated layered temporal responses, with short-term oscillations coexisting alongside persistent long-term synchronization. Institutional processes, particularly budgeting and legislative cycles, contributed to sustained medium- and long-scale coherence, indicating that usage behavior was embedded within broader organizational rhythms. E-resource usage was not random, but systematically shaped by multi-scale external dynamics, with direct implications for collection management and budgeting. The persistent long-term coherence with legislative and budgeting cycles indicated that collection development might be strategically aligned with institutional calendars, rather than driven by short-term demand. The library may anticipate increased demand for both journals and books 6–12 months before major legislative sessions, prioritizing subscriptions, renewals, and access expansions in advance. Similarly, academic libraries might align acquisitions with curriculum cycles and research funding periods, recognizing that demand peaks are structurally embedded and foreseeable. The strong long-term influence of the pandemic and economic crisis suggests that libraries should adopt resilience-oriented and flexible collection strategies. Libraries should maintain a core portfolio of high-demand e-resources, which remain stable even under external shocks, while applying adaptive mechanisms for short-term fluctuations, such as temporary access models, evidence-based, or demand-driven acquisition. The divergence between journal and book usage patterns suggests the need for format-specific management strategies. Journals, which exhibit stronger structural and long-term coherence, should be managed through stable, subscription-based models with strategic continuity, while books—more sensitive to short-term and medium-term fluctuations—are better suited to flexible acquisition models, such as demand-driven acquisition, short-term loans, or usage-triggered purchases.
The integration of wavelet-based decomposition with regularized regression and interpretability tools provides a robust analytical framework capable of distinguishing between transient variability and structural regularity. The findings confirm that accurate interpretation and forecasting of e-resource usage require both sufficient temporal depth and analytical approaches sensitive to scale-dependent dependencies. E-resource usage is highly sensitive to immediate user needs, recent access conditions, and short-term external events, rather than being driven primarily by slow, long-term trends. This requires implementing frequent monitoring and rolling adjustments of acquisition models, based on recent usage patterns. The strong performance of ElasticNet, Ridge, and Lasso models suggest that library usage follow stable and structured relationships, which could be effectively analyzed through regularized approaches. Accordingly, libraries can use short-term forecasts to adjust subscription levels, negotiate flexible licensing agreements, or optimize resource allocation based on expected demand. Journal usage, while highly dynamic, exhibited slightly more structured patterns and benefits from models such as ElasticNet, suggesting the need for stable, but flexible, subscription strategies. In contrast, book usage—better captured by Ridge regression—showed stronger short-term variability, supporting the use of highly flexible acquisition models, such as usage-triggered purchases or short-term loans. The SHAP analysis integrated into the forecasting framework indicated that the most influential predictors were very recent usage components, implying what users accessed in the last one or two months were the strongest determinant of future demand. This suggests that in collection management libraries should adopt decision systems, in which acquisitions, renewals, and cancellations are continuously updated, based on the short-term usage.
Practical implications and future research
The findings extend signal-processing applications in LIS by revealing multi-scale e-resource usage patterns with both structural and event-driven characteristics. Integrating wavelet-based decomposition with econometric modeling can support evidence-based collection planning and budget allocation decisions in academic and parliamentary libraries. Future research may test the generalizability of these multi-scale dynamics across different library types and national contexts, incorporate additional exogenous variables such as pricing policies and user demographics, and explore alternative multiresolution or nonlinear forecasting frameworks to further refine predictive robustness.
Generally, this research indicates that the future of library collection management lies in proactive, cycle-aware, and data-driven strategies, in which libraries function not only as repositories of information, but also as adaptive systems synchronized with broader societal, economic, and institutional rhythms.
Footnotes
Acknowledgments
We sincerely appreciate the editorial input provided by our human native speaker connection.
Declaration of generative ai and ai-assisted technologies in the writing process
During the preparation of this work, the authors used ChatGPT (OpenAI) to improve the language and readability of the manuscript. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.