FIFA has recently allowed the use of electronic performance and tracking systems (EPTS ) in professional football competition, providing teams with novel and more accurate data. Physical performance has not yet taken much attention from the research community, due to the di culty of accessing this information with the same devices during training and competition. This study provides a methodology based on machine learning and statistical methods to relate the physical performance variation of players during time-framed training sessions, and their performance in the following matches. The analysis is carried out over F.C. Barcelona B, season 2015-2016 data, and makes emphasis on exploiting the design characteristics of the structured training methodology implemented within the club. The use of summarized physical variation data has provided a remarkable relation between higher magnitudes of variation in 3-week time frames during training, and higher physical values in the following matches. With increased data availability this and new approaches could provide a new frontier in physical performance analysis. This is, up to our knowledge, the rst study to relate training and matches performance through the same EPTS devices in professional football.
Professional football has attracted the attention of the data science community
in the last decade due to the increasing availability of quantitative data. The
latest technology has provided the possibility of gathering di erent kinds of
speci c metrics, from team statistics to in-game detailed events, contributing to
the improvement of typical and critical tasks such as team tactics evaluation,
opponent analysis, player scouting and training design. The idea that exploiting
data-related analysis can become a competitive advantage within professional
sports is increasingly supported [
At F.C. Barcelona, EPTS devices have been recently used to aid the
evaluation of the applied training methodology, the structured training, a system
that sets the baselines for the planning and adaptation of the training activities
along the season, providing the novelty of incorporating competition activities
in this design. This involves the idea of providing a schema in which the player
is promoted to adapt to the training demands and evolve in each of its
structures, beyond the strictly physical conditions [
The main objective of this study is to nd whether there exist signi cant relations between physical performance of players during training and the measured performance in subsequent matches, for F.C. Barcelona B data from season 2015-2016. Machine learning algorithms are used in order to exploit the contribution of the high amount of measured variables as a whole, all of which are expected to contribute explaining the player's dynamic up to some extent. The study is structured in three main stages. A data preparation stage in which data is pre-processed and normalized, and two datasets are created. An exploration stage where dynamic time warping and cluster analysis is applied in order to obtain representative natural groups from data. And nally, a validation stage, where the matches associated with clustered series are extracted and statistical tests are performed to determine the existence of signi cant di erences. Final conclusions and future work suggestions are detailed, regarding the usefulness of this approach and the nding of moderate standardized di erences between groups presenting high and low variations of physical values from week to week. 2 2.1
F.C. Barcelona B has collected both training and matches physical performance
measurements, for season 2015-2016, using the StatsSports GPS Viper Pod
devices. The resulting tracking information is manually segmented by physical
coaches, and further visualized through a software integrated with the devices
which outputs several variables. From this set of variables, we have selected 15
along physical coaches, described in Table 1, which summarize the considered
most relevant performance information. Variables are structured in three main
groups: locomotor, metabolic and mechanical. Locomotor variables refer to
simple direct measurements of travelled distance and speed, that are obtained solely
through GPS. Metabolic variables are associated with energy expenditure and
exertion, while mechanical variables relate with intensity changes and impacts
[
The dataset is initially processed, adding additional contextualization variables and performing several types of normalizations. Within F.C. Barcelona training structure, training days are labelled in strict relation with the following match day, where match is labelled as MD, the following two days MD+1 and MD+2, and the previous days MD-1 up to MD-4. Each day-type follows speci c design rules for training drills. For simplicity of the study, only day MD-3 sessions are used, due to they similarities to match days in terms of number of players, playing spaces and opposition level. Additionally, day MD-3 involves the highest di erences between physical values. Goalkeepers are deleted from the database since they face considerably di erent physical challenges than eld players. A new variable, load percentage (PER) is added in order to re ect the session load, which is calculated as a ratio of the average AMP from matches. All the measured values are normalized by dividing by the total time of duration of the session. Variables that already represent averages or maximums are kept
Name and
Acronym
Average Metabolic Energy expended by the player per second per kg,
meaPower (AMP) [
Description
Name and
Acronym
Fatigue Index (FAI) Accumulated DSL from the total session volume, in
[
Each value Vi represents the absolute di erence between a value registered at sessions Si+1 and Si. Two datasets were built: the rst one consists of timeseries of W window size. A sliding window approach is followed by using a xsized (W ) window of consecutive weeks. The time-series dataset is conformed by groups of W rows containing the 15 physical variables, corresponding to a player in a speci c period of the season. Selected windows sizes during experiments are 3 and 6 in order to match the methodology of the club. Windows are moved SW steps each time, so to control the degree of coincidence of values between windows. The value of SW was selected following Equation (1) to avoid an excessive overlap between windows and to avoid a too strict separation that would reduce signi cantly the amount of data. Another dataset is built which summarizes each group of W rows in each variable, by calculating the average of absolute di erences. Equation (2) describes the performed calculations, where Pjvd corresponds to the absolute average of window di erences of a variable v of a player j, measured in the window frame d, substracted by the mean of Pivd for every other player i. P corresponds to the set of all possible players.
SW = W
(W=3) Pjvd =
PiW=+21kSi
Si 1k W
PjiP6=jj Pivd jP j (1) (2) 2.3
Visual Exploration. Speci c di erences of physical variables where assessed
visually through boxplots and analytically through one-way ANOVA and Post
Hoc tests observing the di erences between type-days (i.e MD-4, MD-3, MD-2,
etc.). A PCA analysis was also performed, and the two principal components
where plotted accounting for 69% of variance and observing the acknowledged
di erences. On the other hand, di erent plots over the time-series and
summarized datasets allowed to visualize oscillatory patterns along the season that
respond to cycles design. Also, it is observed how players tend to oscillate in
similar patterns due to the training design. There exist, however, several cases
in which certain players magnitude of variations starts di ering considerably
from the mean variation. The results of these observations coincided with the
understanding of physical responses in training from the club's physical coaches.
For space restriction reasons, the graphical results are omitted from this section.
Calculating Series Similarities through Dynamic Time Warping.
Dynamic time warping (DTW ) is a highly used method that allows to measure the
similarity between two temporal series, while being less sensitive to signal
transformations such as shifting, uniform amplitude scaling or uniform time scaling
[
Results for both the time-series and the summarized datasets are presented
together since they follow an identical approach in its evaluation. For both cases,
the selected number of clusters was 2 by four of the ve di erent indices, the
sample size of the training sessions dataset is 112, and the sample size of
associated matches dataset is 82. Only the results for 3-week window are presented,
since no statistically signi cant relation was found with 6-week window frames.
For each of the variables conforming the two groups (in each dataset) the
standardized di erence of means was calculated to describe the e ect size. The limits
of the e ect sizes are those suggested by Hopkins [
Detailed results are presented in Table 2. It can be clearly observed that for the summarized dataset almost every variable in training registered a moderate to large e ect size when comparing groups. So, we are observing the detection of two groups: one where the average magnitude of variations of each variable is higher (high variation group), and one where is lower (low variation group). It is critical to observe that separation among groups is not absolute, and there exist ranges of values which overlap. This has to do with multivariate nature of the clustering procedure, and coincides with the original expectation of this study. It can also be observed that for the timeseries dataset few variables where able to stand out just with a small size e ect. Even with the selection of Euclidean distance to favor magnitudes, the cluster analysis over the DTW procedure was not able to found a clear separation between groups. The procedure over the summarized dataset, instead, did nd a considerably separation between training groups so the analysis over associated matches is easier to interpret and translate to practice. Figure 2 presents the e ect sizes for the associated matches in both datasets. It can be observed for both cases that variables registering high intensity e orts, energy consumption and distance travelled appear with higher magnitude in the high variation group consistently, while the total load percentage and training minutes in the previous three weeks are considerably low in this same group. HML, AMP and DIS present moderate e ect size in the summarized dataset, variables belonging to metabolic group (the rst two) and locomotor group. For the timeseries dataset only HML presents a moderate e ect size, toward the same tendency. A small e ect size is also observed in other locomotor (MAX), metabolic (PER and HSR) and mechanical variables (DSL, DEC) toward the same tendency. Three-weekly PER and training minutes show also a moderate e ect in di erences, towards lower values. Sample size for associated matches allows to conclude with certainty about moderate size e ects. Small e ects should be taken into account, but must be further validated with the future increase of availability of data. means SDM is presented for each case. Training results refer to the absolute average of variation while matches results refer to the actual measured physical values.
Training (mean SD)
Matches (mean SD) Variable DSL p/m ACC p/m DEC p/m SPR p/m HSR p/m AMP HML HEF p/m
FAI DIS p/m TLO p/m MAX STE
PER 3W Training PER 3W Training Minutes 3W Total PER 3W Average FAI during training re ected in higher values in 11 of the 15 analyzed variables for locomotor (4/4), metabolic (4/4) and mechanical (3/7) groups in the next matches, and also lower training minutes and accumulated load during training. This approach might provide a way for analyzing the adaptation of players to training dynamics, and even to evaluate training design. The procedure follows a series of simpli cations such as the selection of day-type MD-3 which might incur in loss of information. However, this type of calculations can be easily integrated to daily routine performance analysis carried out by physical coaches, without the need of additional systems or requiring high processing times. The ndings provide su cient evidence to suggest the incorporation of this calculation in daily analysis and track its evolution in order to further measure is e ectiveness on relating with match performance.
The summarized dataset allowed a more representative grouping and more conclusive results. In practice, high and low variations can be found directly by using the ranges found by the clustering procedure for each variable. Also, time-window aggregated information is showing to add value for performance analysis and should be considered in future research. On the other hand, DTW could not provide su ciently clear results in this study, most probably due the the short-size characteristics of the analyzed time series and that exact match of variation patterns might be too strict for the few data available. Also, the player normalization seems to favor a cleaner comparison between players, instead of using absolute values which could lead to di erences that are more related to physical characteristics than actual adaptation patterns.
This is the rst study, up to our knowledge, to relate training and match physical values directly registered from player using EPTS devices during training and matches for a whole season. In the following years, with higher availability of data these remarks must be further validated. Future work should incorporate new day-types in the analysis and factors beyond the physical such as tactical information and variables related with psychological information such as the rate of perceived exertion (RPE). The yearly knowledge of physical evolution of training dynamics and even speci c players might provide new insights about the physical preparation of teams and the performance during competition.