Datasets:

SouthernCrossAI
/

Arxiv

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 132k rows

Filename stringlengths 22 64	Paragraph stringlengths 8 5.57k
Processed_Generation_and_annihilation_of_three_dimensional_m.txt	Figure 20. The figure illustrates the topological details of spontaneously generated new nulls near the radial null of the radial-spiral-pair-2 pair at t = 274.91s. These new nulls are generated in a pair and are marked by the arrows. The fan field lines (in yellow) of the radial null of the radial-spiral-pair-1 pair are directed toward the null point, resulting in a topological degree of +1. Meanwhile, the fan field lines (in green) of the spiral null of the spontaneously generated new pair are directed away from the null point, resulting in a topological degree of −1. Lastly, the fan field lines (in red) of the radial null of the spontaneously generated pair are directed toward the null point, making the topological degree +1. Hence, spontaneous generation occurs in a pair, and the annihilation of the radial null of the radial-spiral-pair-1 pair is with the spiral null of the generated pair. Therefore, the net topological degree of the system remains unaffected by the generation and annihilation of 3D nulls.
Processed_Generation_and_annihilation_of_three_dimensional_m.txt	Figure 21. Magnetic field lines are traced over a time span of t ∈ 274.912, 280.720 s and advected with plasma flow. This time span is selected to investigate the dynamics of field lines responsible for annihilation. Four selected field lines, two green and two yellow, are drawn to illustrate magnetic reconnection. At t = 274.912 seconds, the two green field lines are part of the spine and fan plane of the spiral null, connecting from region b to region a (panel (a)), while the two yellow field lines are part of the upper and lower spine along with the fan plane of the radial null, connecting regions c to d and b. As the evolution progresses, the green field lines change their connectivity from region b to region a to region b to e. Similarly, one yellow field line also changes its connectivity from region c to d to region c to b (panel (b)). With further evolution, one yellow field line changes its connectivity from region c to b to region e′ (panel (c)), and then from region e′ to region b to region a′ to region b. These changes in connectivity occur through magnetic reconnection, ultimately resulting in the annihilation of nulls in a pair.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Abstract—We consider the viability of a modularised mecha- nistic online machine learning framework to learn signals in low- frequency ﬁnancial time series data. The framework is proved on daily sampled closing time-series data from JSE equity markets. The input patterns are vectors of pre-processed sequences of daily, weekly and monthly or quarterly sampled feature changes. The data processing is split into a batch processed step where features are learnt using a stacked autoencoder via unsupervised learning, and then both batch and online supervised learning are carried out using these learnt features, with the output being a point prediction of measured time-series feature ﬂuctuations. Weight initializations are implemented with restricted Boltzmann machine pre-training, and variance based initializations. His- torical simulations are then run using an online feedforward neural network initialised with the weights from the batch training and validation step. The validity of results are considered under a rigorous assessment of backtest overﬁtting using both combinatorially symmetrical cross validation and probabilistic and deﬂated Sharpe ratios. Results are used to develop a view on the phenomenology of ﬁnancial markets and the value of complex historical data-analysis for trading under the unstable adaptive dynamics that characterise ﬁnancial markets.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Technical analysis is a ﬁnancial analytical practice that makes use of past price data in order to identify market states and forecast future price movements based on past movements. The techniques typically rely on past market data (price and volume), rather than company assessments using fundamental analysis. We explore the idea that technical analysis has merit in exposing market inefﬁciencies when they are signiﬁed by repeated feature time-series patterns [1], [2].
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	non-random patterns and internal organisation on different averaging scales. Two key price generation processes have emerged: the low-frequency domain (the result of sequences of closing auctions generating prices), and the high-frequency intra-day domain driven by order-ﬂow itself. Here we consider low-frequency daily sampled data that is the result of the price discovery from closing auctions.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Even at low-frequency, identifying patterns and structure is simultaneously reasonable and notoriously difﬁcult. While it is often clear in hindsight that patterns exist, the amount of noise and non-linearity in the system can make prediction challenging. Fittingly, neural networks are a popular choices for modelling within ﬁnancial markets because of their ability to perform well as universal approximators [5].
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Practical approaches to money management within the realities of adapting and changing market systems increasingly favour online methods, in particular [6] explored the appli- cation of online learning models in this space in the South African market to show that direct (but simplistic) online pattern-learning is able to identify and potentially exploit trading opportunities on the JSE through the assessment of Open High Low Close (OHLC) data. This was extended by [7] to more directly explore the use of online learning ap- plied to optimizing parameters for traditional technical trading indicators as applied to maximising wealth trading zero-cost portfolio strategies.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The work presented here ﬁts into the growing body of work which considers mechanistic and brute-force approaches of ap- plying machine learning models to ﬁnancial market data. The complexity, non-linearity, noise and stability of ﬁnancial mar- kets are highlighted through both the successes and challenges found in training these models. These difﬁcult dynamics, and their notable difference when compared to other popular areas of ML research - which are often around Independently and Identically Distributed (IID) datasets - present fundamental problems to be explored; both in terms of prediction efﬁcacy as well as validation. We present a framework using batch ofﬂine and online learning on JSE closing data, feature extraction and robust non-parametric validation techniques.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Financial academic literature is currently facing a problem in terms of validation and veriﬁcation of results. Trading strategy proﬁtability has typically been proven using historical simulations, or “backtests”. However, the recent advances in technology and algorithms available to construct these strate- gies have resulted in researchers being able to test increasing numbers of variations of factors. This has made it increasingly difﬁcult to control for spurious results. The problem is so extensive that some meta-research papers suggest that most published research ﬁndings are false [8].
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The standard way of implementing backtests is to split the data into two portions: an In Sample (IS) portion which is used to train the model, and an Out of Sample (OOS) portion which is used to test the model and validate results. If vast numbers of different model conﬁgurations are tested, then it is only a matter of time before false positives occurs with high performance both IS and OOS (i.e. overﬁtting) [9], [10]. The nature of ﬁnancial data makes it difﬁcult to resolve these issues effectively. There is a low signal-to-noise ratio in a dynamic and adaptive system with only one true data sequence. Traditional hypothesis testing frameworks (e.g. Neyman-Pearson) are not sufﬁcient in this context making more sophisticated techniques necessary.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The problem of overﬁtting is not novel. However, in a machine learning context, frameworks are often not suited to trading schedules with a random frequency structure. They do not account for overﬁtting outside of the output parameters nor take into consideration the number of trials attempted. The common ‘hold-out’ strategy is where a certain portion of the dataset is reserved for testing true OOS performance. Numerous problems have been pointed out with this approach. The data is often used regardless, and awareness of the movements in the data may inﬂuence strategy and test design [11]. For small samples, a hold-out strategy may be too short to be conclusive [12]. Even for large samples, it results in the most recent data (which is arguably the most pertinent) not being used for model selection [9], [13]. We present a novel application of existing sophisticated validation methods (see Section III) to a machine learning framework.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	ideas to- gether: i.) SAE based feature selection, ii.) Deep Learning with pre-training and weight initialization, and iii.) Online Learning and Backtest Overﬁtting Validation. The learning part of the framework consists of two phases: I. batch learning and II. online learning.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	I Batch learning phase: IS data is used to train Stacked Autoencoder (SAE) networks which in turn are used to perform feature reduction for Feedforward Neural Net- works (FNNs) learning the price ﬂuctuation predictions in the IS data. Both are trained using Stochastic Gradient Descent (SGD).
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	II Online learning phase: The batch trained FFN networks are used to predict price ﬂuctuations on OOS data through Online Gradient Descent (OGD).
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	These online predictions are then sequentially used to simulate trading in a Money Management System (MMS), which in turn generates simulated returns. Finally, the MMS returns are used as input for the Probability of Backtest Overﬁtting (PBO) and Deﬂated Sharpe Ratio (DSR) techniques in order to validate the legitimacy of the framework.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The process has two key principles: First, implementing a generalised version of a system which could offer exploration of more complex techniques. Second, ensuring an effective modularisation of steps such that the process can be recon- ﬁgured accordingly while maintaining its integrity. In doing so, a separable system is created which brings together all key concepts. We aim to delivery the simplest implementation of a complex framework such that the effects of individual components can be properly assessed and developed. The full process ﬂow can be seen in Figure 7 in the Appendix.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The aggregations are scaled using a modiﬁed Normalization, where the min and max values are determined by the Training portion of the dataset, but applied to both the Training and Prediction portions. This emulates a production implemen- tation where future data is unknown. The log-differenced, aggregated and scaled data is then used as the input for the neural network models. Predicted outputs have the scaling and log differencing reversed in order to reconstruct the actual price point predictions for performance assessment.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The problem of vanishing and exploding gradients has been one of the primary barriers to deep learning with neural networks. The approach of greedy layer wise pre-training for SAEs was suggested by [14], which allowed much deeper layered networks to be trained than previously possible [15]. Once the SAE is trained, a ﬁnal output layer is added and the entire network can then be ﬁne-tuned through back- propagation without suffering such performance degradation from vanishing or exploding gradients, enabling training of both SAE and FNN networks [16], [17].
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	In Section V we see that batch training on historical data has a limited beneﬁt, which gives primacy to weight initialization techniques for machine learning of ﬁnancial time series. Initial results found that RBM pre-training for sigmoid SAE networks (as described by [17]) had detrimental effects on network performance. This suggests that the IID assumptions and the different loss functions result in the ﬁnancial time series data used being pathalogical for RBM pre-training, and is discussed further in [18]. It has also been shown that pre-training may largely act as a prior which may not be necessary if large enough datasets are available [19], [20]. In the context of ﬁnancial time series this prior can explain the poor perfor- mance. For these reasons we focus on variance based weight initialisations developed by [21] and [22]. These have simpler implementations, faster computation and enables initialization for non-probabilistic activation functions, such as ReLU.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The beneﬁt of the modularised system is emphasised here, as the SAE training will not suffer from limitations due to backtesting considerations: any amount of conﬁgurations can be tested for feature extraction without concern. The best chosen SAE networks (based on a minimum Mean Squared Error (MSE) score) are used to reprocess both the Training and Prediction datasets such that the input is encoded, and the output is as before. These encoded datasets can then be used for the subsequent steps in the framework. We did not implement a step to update the SAE, but results detailed in Section V suggest that this would be an important inclusion in a production system.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Once the predictive network is trained on IS data using SGD, the OGD process is run through the encoded Prediction dataset in order to generate the predictions for the asset prices that the model produces - thus emulating what would have occurred in a live environment.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	trading strategy. It is important, in the interest of effective optimization, that the pattern prediction of the system is not tightly coupled with making it proﬁtable. Thus, the modularity of the system is continued with a separation between the prediction signal and the MMS implementation.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Validation is implemented with Combinatorially Symmetric Cross-Validation (CSCV) [9] that uses the IS and OOS returns from the MMS, which in turn uses the prices from the prediction network; which is a somewhat novel application. Conceptually, the whole system comes into place here, as the results from the CSCV process are now indicative of not only backtest overﬁtting in the trading strategy, but also in the prediction network and without having to consider the impact of the many conﬁguration tests for feature extraction. A modiﬁed version of ONC was run, with reduced cluster exploration. As noted in [18], this did not appear to affect results.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	CSCV was developed by [9] as a robust approach to assess- ing backtest overﬁtting. Their research deﬁnes backtest over- ﬁtting as having occurred when the strategy selection which maximizes IS performance systematically underperforms the median OOS performance in comparison to the remaining conﬁgurations. They use this deﬁnition to develop a framework which measures the probability of such an event occurring, where the sample space is the combined pairs of IS and OOS trading performance measures. The PBO is then established as the likelihood of a conﬁguration underperforming the median OOS while outperforming IS.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The CSCV methodology is generic, model-free and non- parametric, allowing it to arguably be used in any model case. By recombining the slices of available data, both the training and testing sets are of equal size (advantageous for comparing performance statistics). The symmetry of the set combinations in CSCV ensure that performance degradation is only as a result of overﬁtting, and not arbitrary differences in data sets. There is no requirement of a hold-out set, which removes potential credibility issues regarding whether the holdout set was treated appropriately or not. The logit distribution developed through the assessment offers a useful view on the robustness of the strategies used and the nature of the PBO score.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The PBO can be estimated using the CSCV method results, which provides an estimate rate at which the best IS strategies underperform the median of OOS trials. [9] extend this to show that with models overﬁtting to backtest data noise, there comes a point where seeking increased IS performance is detrimental to the goal of improving OOS performance.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	the true SR is positive under consideration of numerous trials being tested [24]. The DSR can be estimated using the PSR methodology as (cid:91)P SR[SR∗] where the benchmark Sharpe ratio, SR∗, is calculated using the False Strategy Theorem. The calculation of SR∗ requires both the variance of trial SR values and the number of independent trials, which are not typically consid- ered and where determining independence is challenging. [25] aim to resolve this with the Optimal Number Clusters (ONC) algorithm, a modiﬁed K-means methodology of clustering strategies and trial results. This clustering allows an estimation then of both the variance and number of trials, which in turn allows the DSR to be calculated. With this as a conﬁdence level, one can accept or reject the notion that the observed (cid:99)SR is positive.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Datasets were constructed using JSE closing price relative data for 2003-2018 [26], with a 60:40 split on the Train- ing:Prediction subsets. The closing price dataset consisted of 10 assets from the JSE Top 40: AGL, BIL, IMP, FSR, SBK, REM, INP, SNH, MTN, DDT (coming from a variety of sectors). More source information, data snapshots and price charts are available in [18].
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The software libraries, written in Julia, were produced for all the training, experimentation and recording of results. These are discussed extensively in [18], and made available online [27].
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The parameter space is explored using a phased grid search approach. For each stage, the relevant parameters are each speciﬁed as a set of values, and all sets are then used to generate the full combinatorial space, such that each possible combination of the speciﬁed parameters is tested.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	We expected that the IS batch training using SGD for the predictive network would improve OOS P&L performance. Theoretically, the training on historical data might prime the network for predicting future data. However, we found that the effects of IS training had limited beneﬁt. We ran experimental trials to test the hypothesis that the amount of historical IS data available is of limited beneﬁt. We found the P&L results validate this idea, as seen in Figures 1 and 2. We saw that extensive training on past data may be akin to pre-training network weights at best, and counterproductive in overﬁtting to dynamics that no longer exist at worst. This highlights the complexity and dynamic nature of ﬁnancial time series, where past relations and behaviours are not necessarily indicative of present state. It follows that the primary determinants of OOS P&L are those present in the OGD (OOS) learning phase: the OGD learning rate, the data horizon aggregations, and the SAE feature selection.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	This ﬁts well with research showing that online algorithms typically perform as fast as batch algorithms during the ‘search’ phase of parameter optimization, but that ‘ﬁnal’ phase convergence tended to ﬂuctuate around the optima due to the noise present in single sample gradients [28], [29]. [30] showed that to consider the is actually more practical convergence towards the parameters of the optima, rather than the optima itself (as deﬁned by the cost function) - the difference between the learning speed and optimization speed, respectively. Online learning methods are thus well suited to ﬁnancial market modelling using neural networks. They allow effective and efﬁcient incremental updates as more recent (and relevant) data becomes available. Further, the increased learning speed over optima convergence makes them a ﬁtting choice when data is non-IID and constantly changing.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Fig. 1. These results show OOS P&L grouped by the number of epochs in the SGD IS training phase. Here 100 Epochs offered the best overall performance, and further training to 500 or 1000 epochs degraded performance due to the network overﬁtting on the IS data. The results show that the beneﬁt of historical data is limited - having networks become better at learning return relationships from 10 years in the past did not lead to increased OOS P&L for more current data. The small difference in the upper half of observations between 10 and 100 Epochs further emphasises this point.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Fig. 2. To further explore the effect of IS training on historical data, conﬁgurations were run with a percentage of the usual training data excluded, with the P&L results grouped above. The exclusion of up to 80% of the IS training data resulted in only a 2.2% drop in median OOS P&L for those networks. The training in these instances was not adjusted to increase the number of epochs according to the size of IS data, and so the conﬁgurations with more data excluded were also in essence trained less.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	resulted in easier replication; while longer horizons are more difﬁcult (as indicated through MSE scores). The reproduction differences are discussed in [18].
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	the SAE feature size decreased from 25 to 101, SAEs learnt longer term features (as they were increasingly unable to represent short term ﬂuctuations). For FNN networks with larger learning rates which could otherwise adapt quickly, the increased focus on long term features caused P&L perfor- mance degradation. For FNN networks with smaller learning rates, poorly able to adapt quickly to new information either way, there was a beneﬁt from SAE features with an increasing representation of long term trends. The relationship is empha- sised dramatically in the 10-feature SAEs, to the point that lower learning rates can be more effective in generating OOS P&L. The P&L performance suggests that the 10-feature SAEs overﬁt to long term IS features, and became pathological for short term adaptation OOS.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The most noteworthy results were the 5-feature SAEs, where performance was often on par or better than the higher feature sizes, or no SAE at all, as seen in Figure 3. It is possible that the small encoding layer acts as a form of regularization, forcing the SAE to learn more consistently generalisable features. The performance of the 5-feature networks, with an 83% reduction in input data, is clear evidence of the efﬁcacy and potential of feature selection in ﬁnancial times series.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The effect of data horizon aggregations is as expected: short term horizons (i.e. [1,5,10]) outperformed in conﬁgurations with more SAE features and higher learning rates; longer term horizons (i.e. [10, 20, 60]) outperformed in low learning rate and low feature conﬁgurations. The differentiation between these groups is seen more robustly in Section V-F, where data horizon aggregations are determined to be the primary clustering attribute for trade correlations. Strategies focusing on short term predictive strategies (aggregations of [1, 5, 10]) had higher variance in returns than the longer data horizon strategies, though also the highest highest P&L and Sharpe ratios. This again shows the beneﬁts in focusing on recent cross-sectional data in ﬁnancial markets. The differentiation between the groups is discussed more in [18].
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The benchmark is an upper bound on performance, repre- senting MMS returns based on perfect knowledge of future prices. The benchmark full return rate is 2.4% with trading costs, over a period of 1555 trading days. So while the strate- gies’ proximity to the benchmark do represent a framework success, they are not necessarily representative of a feasible market solution. Ultimately, this enforces the notion that the MMS implementation is of exceeding importance in a live trading process, and predictive accuracy is only able to achieve so much.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Figure 4 shows the distributions of OOS P&L with trading costs being accounted for. There were a signiﬁcant number of conﬁgurations within 20%-30% of the benchmark. The trials with 0 P&L are networks which suffered from either exploding or vanishing gradients, and were not able to make sufﬁcient predictions.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Fig. 3. This ﬁgure shows that the lower learning rates (0.005, 0.01) performed best with strategies using long term trend pricing. The 10 feature encoding appeared to optimise speciﬁcally for this perspective. The optimisation caused outperformance at the lower learning rates and detrimental performance at higher learning rates (which perform best with short term ﬂuctuation strategies). The 15 to 25 horizon encodings showed a better association to the short term strategies. Here higher encodings and higher learning rates offer the best performance. The 5 feature encoding offered consistent performance across learning rates and shows the learning of generalisable features.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Fig. 4. The distributions of all OOS P&L values, with the benchmark P&L indicated in orange, show an encouraging view of the results. There is a signiﬁcant negative skew, with a proportionally small number of strategies resulting in negative returns, even with capital and trading costs applied. There were a large proportion of strategies near the OOS upper bound, which is within 28% of the benchmark.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	It is noted that the logit metric, which the CSCV method relies on, has its basis in an ordinal ranking; indicating whether the best strategy in the IS set is higher than the median in the OOS set. This means that poor performing conﬁgurations can artiﬁcially bolster an ordinal position past the median point and so bias PBO results. An honest, wide exploration of the parameter solution space in a mechanistic machine learning framework is likely to result in “poor” conﬁgurations being tested (as visible in the ‘0’ P&L conﬁgurations in Figure 4). As a result, the methodology shifts the onus onto the researcher in both handling and reporting these dynamics responsibly.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Further to this, the parameter space search methodology (Section IV-B) may also result in a lower likelihood of PBO due to the way of combining parameters across IS and OOS stages. By way of example, any conﬁguration which performs well IS will have all possible OOS parameters tested in combination with it. While some of these combinations may result in poor performance, there will always be a combination of the best IS and best OOS parameter choices. This makes it unlikely that the best conﬁgurations will be past the median point for the logit calculation, resulting in a systematically low PBO regardless of how many conﬁgurations are attempted.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Lastly, the CSCV algorithm requires a parameter choice of how many windows to split the data into. While not inherently problematic, impact on this choice can have a signiﬁcant results which is not visible in the reported PBO value. We discuss this further in [18].
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Fig. 5. The CSCV logit distribution for the 21,653 conﬁgurations run, with a calculated PBO of 1.7%. The strong negative skew is indicative of IS and OOS strategy returns being closely matched in rank and results in a low PBO score. This is a favourable assessment for the efﬁcacy of the full framework presented here and shows that training was able to occur without much risk of backtest overﬁtting.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The ONC algorithm produced three clusters: one consisting mostly of the negative Sharpe ratio conﬁgurations, and two containing the remaining conﬁgurations partitioned by their data horizon aggregations conﬁgurations. The distributions for all three clusters’ Sharpe ratios can be see below in Figure 6. If we consider the two primary clusters, we see that Cluster One contained all the trials with horizon aggregations of [5, 20, 60] and [10, 20, 60], and Cluster Two contained all trials with horizon aggregations of [1, 5, 10]. The nature of the experimentation process, with the combinatorial parameter space exploration (as detailed in Section IV-B), is such that other parameters were mostly evenly split across the 2 clusters (e.g. OGD learning rate, network sizes, initializations and so on).
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The clusters here indicate that the networks adapted to at least two different general strategies for predicting prices: one which which was more inﬂuenced by the long term ﬂuctuations, and the second was more inﬂuenced by the short term ﬂuctuations. The results presented in Section V-B are then indicative of the networks ability to execute these overarching strategies effectively.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The best Sharpe ratio value (with trading costs applied) was 0.64 and part of Cluster Two, with the [1, 5, 10] price ﬂuctuation horizon aggregations. The distributions seen in Figure 6 indicate that at a general level, Cluster One has more consistent performance, Cluster Two on the other hand has higher variance, with more strategies at both the lower and higher range of Sharpe ratios. The lack of further clusterings was probed manually to ﬁnd that the variance from further subclustering lead to a worse cost function score.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Fig. 6. This ﬁgure shows the distributions of all Sharpe ratios, grouped by the clusters produced by the ONC algorithm, and an indication for the best Sharpe ratio (which is in Cluster Two). Cluster One has more consistent values, with a higher mean (µc1 = 0.393, µc2 = 0.375) and higher negative skewness (˜µ3,c1 = −0736., ˜µ3,c2 = −0.543). Cluster Two has higher variance (σc1 = 0.022, σc2 = 0.028) but with more strategies at both the lower and higher range of Sharpe ratios; including the highest Sharpe ratio from all trials.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	variance estimates to be used to calculate SR∗ (the maximum expected observed Sharpe ratio due to variance under the null hypothesis of H0 : (cid:99)SR = 0). Using SR∗ as the benchmark, the PSR calculation ((cid:91)P SR[SR∗]) can then be used to determine if the observed (cid:99)SR is a false positive or not. This gives us the DSR as a conﬁdence for observing a positive best SR. The benchmark SR∗ calculated was 0.211245, and the best (cid:99)SR observed was 0.642632, leading to a (cid:91)P SR[SR∗] of 1.0, thus indicating that the trials certainly contain a strategy which has a positive SR rate. This seems a reasonable conclusion, considering the SR distributions in Figure 6.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Mechanistic machine learning approaches to ﬁnancial mar- ket data hold some promise for enhancing the performance of low-frequency quantitative trading. To investigate this potential we provide a novel framework that we show to be effective in both training, and in validation. The framework is con- ﬁgurable and based on decoupled modules and uses several well understood techniques: deep-learning neural networks for stock price ﬂuctuation prediction, stacked autoencoders for the purpose of feature selection, both CSCV & PBO to assess the returns from MMS, and the likelihood that backtest overﬁtting has taken place, and DSR in order to assess the likelihood of a positive Sharpe ratio having been observed.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	markets over time need careful attention. Learning optimiza- tions for IS training, such as regularization and learning rate schedules, were shown to have IS beneﬁts, but little impact on OOS performance [18]. This gives weight to the importance of online learning methods for ﬁnancial applications, and in turn highlights the importance of network initialization. The results showed better performance for the He-adjusted initialization and poor performance in RBM based pre-training.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The primary determinants of OOS P&L were shown to be those which affect the online learning data and the model’s ability to adapt to this. We found SAE encoding layer sizes inﬂuenced the nature of features learnt, with smaller encodings generally leading to learning of longer term features. This relationship continued from layer sizes 25 to 10, with increas- ing effect. The results from the 10-feature SAEs suggest the SAEs were overﬁt to long term IS features and pathological for OOS adaptation. The relationship changed at 5-feature SAEs, which learnt far more generalisable features. The 5-feature SAEs had very competitive performance and show that feature selection in ﬁnancial time series is both possible and beneﬁcial despite the complexity present. Predictive strategies focusing on long term changes were present in conﬁgurations with longer data horizons, less features and lower learning rates (slow adaptation). Short term strategies presented with shorter data horizons, more features and larger learning rates (quick adaptation). The data horizon was the primary separating attribute in the ONC clusters, emphasising these groupings. The short term strategies had higher variance, but also the highest returns. This again highlights both the increased value in recent information in ﬁnancial markets, as well as the difﬁculty in using it due to the amount of noise present.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	The most challenging aspect of a mechanistic approach to learning is avoiding backtest overﬁtting (see Section I-B). Probing and validating of the generalisation error was done using the PBO methodology in conjunction with DSR [25]. The results discussed in Section V show a low likelihood of the models having overﬁt. The CSCV and PBO methodologies suggest that they are able to add-value to our novel implemen- tation of machine learning models; while providing a robust assessment of results. The results were further validated using the ONC and DSR algorithms to detect positive Sharpe ratios. The phenomenological view of ﬁnancial markets based on our experimental results suggest a very limited beneﬁt to training on long term historical ﬁnancial time series. A in cross sectional view of the data has far more weight delivering OOS returns; this is noteworthy in the context of neural networks. Based on our simulation work we speculate that money management strategies can be more important determinants of OOS proﬁtability relative to signal generation and should also be learnt.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	https://zivahub.uct.ac.za/articles/dataset/JSE Top40 Closing Price Relative Data for 2003-2018/11897628, 2020. Accessed: 2020-06-30. [27] J. Da Costa, “Julia libraries for online non-linear prediction of ﬁnancial time series patterns.” https://github.com/joel11/Masters, 2020. Accessed: 2020-06-30.
Processed_Learning_low-frequency_temporal_patterns_for_quant.txt	Fig. 7. Overall Process Flow: The ﬂow diagram here offers a visual representation of the framework from data processing through to training and price prediction processing, as detailed in Section II. The PBO & DSR validation steps are not shown here, though use the IS and OOS Price Predictions from step 8 as inputs.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Quantum Chromodynamics exhibits a hadronic conﬁned phase at low to mod- erate temperatures and, at a critical temperature TC, undergoes a transition to a deconﬁned phase known as the quark-gluon plasma. The nature of this decon- ﬁnement phase transition is probed through visualisations of the Polyakov loop, a gauge independent order parameter. We produce visualisations that provide novel insights into the structure and evolution of center clusters. Using the HMC algorithm the percolation during the deconﬁnement transition is observed. Us- ing 3D rendering of the phase and magnitude of the Polyakov loop, the fractal structure and correlations are examined. The evolution of the center clusters as the gauge ﬁelds thermalise from below the critical temperature to above it are also exposed. We observe deconﬁnement proceeding through a competition for the dominance of a particular center phase. We use stout-link smearing to remove small-scale noise in order to observe the large-scale evolution of the center clusters. A correlation between the magnitude of the Polyakov loop and the proximity of its phase to one of the center phases of SU(3) is evident in the visualisations.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Quantum Chromodynamics (QCD) is the gauge ﬁeld theory that describes the strong interactions of quarks, the constituent particles of hadrons such as protons and neutrons. In QCD, the strong interaction is mediated by a gauge boson known as the gluon. The self-coupling of gluons through the color charge gives rise to a non-trivial vacuum structure, conﬁning quarks and generating mass through dynamical chiral symmetry breaking.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	To observe this phase transition in Lattice QCD simulations, one examines the behavior of a complex-valued observable known as the Polyakov loop which acts as an order parameter. It has an expectation value of zero in the conﬁned phase and a nonzero expectation value in the deconﬁned phase [9]. As we will observe, this transition occurs through the growth of center clusters, regions of space where the Polyakov loop prefers a single complex phase associated with the center of SU(3). It is these clusters that we analyse in this paper.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	We will demonstrate deconﬁnement in the behavior of the Polyakov loops at TC and produce and analyse visualisations of the center clusters that allow the evolution of the center clusters to be directly observed for the ﬁrst time. We explore both the HMC evolution and the percolation as T is increased above TC and establish a correlation between the peaks in the magnitude ρ((cid:126)x) and the proximity of φ((cid:126)x) to a center phase. This conﬁrms the underlying assumption of Ref. [5], which links the center domain walls to phenomena observed at RHIC and the LHC. This allows for a better understanding of the nature of the phase transition in vacuum. To do this, we develop a custom volume renderer that can correctly visualise the 3D complex ﬁeld of local Polyakov loops.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	In the conﬁned phase, the free energy of a static quark is inﬁnite, so = = 0. In the deconﬁned phase, the free energy of a static quark is ﬁnite, P (cid:105) (cid:104) so = 0 [10]. Thus, the Polyakov loop is an order parameter for = L((cid:126)x) (cid:105) (cid:104) conﬁnement.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	As it is deﬁned here, the Polyakov loop would be exactly zero in both phases, as in pure SU (N ) gauge theory, all N sectors are equally weighted in the par- tition function. Therefore usually the absolute value of the Polyakov loop is used as an order parameter. An alternative approach, which we take, is to re- move this remaining symmetry by performing center transformations to bring the most dominant center sector in each conﬁguration to the same phase. In full QCD the fermion determinant introduces a preferred phase, causing the peak with a phase of zero to always become dominant above the critical temperature [10], so this would no longer be a concern.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	In addition, on the lattice the expectation value of the absolute value of the Polyakov loop below the critical temperature is not exactly zero, due to ﬁnite volume eﬀects. However, the volumes we work with are large enough that it is close to zero. In the thermodynamic limit, the Polyakov loop truly vanishes below the critical temperature.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Figure 1: Histograms of the distribution of complex phases of local loops across a gauge ﬁeld ensemble, showing diﬀering occupation of the three center phases.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	so if the center symmetry is conserved, = 0 [16]. Thus, the Polyakov loop L((cid:126)x) (cid:105) (cid:104) is an order parameter for the center symmetry and deconﬁnement corresponds to a spontaneous breaking of the center symmetry.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	On the lattice, the temperature T is related to the temporal extent aNt by T = 1 . The volume V = (aNs)3 depends on the spatial extent of the aNt lattice aNs. Most studies of center domains change the temperature by varying the lattice spacing [9–11, 17]. This is a problem because the volume of the lattice is changed as the temperature changes. The only way one can change the temperature of an isotropic lattice without changing the volume is by adding or removing lattice sites in the time direction. In order to be able to adjust the temperature in a continuous manner, we instead introduce anisotropy into the lattice, replacing the single lattice spacing a with a spatial lattice spacing as and a temporal lattice spacing at. In this way we can vary the temperature T = 1 while maintaining a constant volume V = (asNs)3 by varying at whilst holding as ﬁxed. This allows us to observe the evolution of the center clusters in gauge ﬁeld conﬁgurations moving from a conﬁned conﬁguration to a deconﬁned conﬁguration under the HMC algorithm with no change in physical volume. This cannot be done with any of the previous methods and is the ﬁrst such presentation.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	with probability distribution ρ(U ) = e−SG[U ]. To do this, it is common to use local updates such as the pseudo heat-bath combined with over-relaxation algorithms as this is an eﬃcient way to generate pure gauge ﬁeld conﬁgurations. However, we are ultimately interested in dynamical fermion results, so we use the Hybrid Monte Carlo (HMC) algorithm [19]. This allows us to study the way the HMC algorithm evolves the gauge ﬁelds at ﬁxed temperature and from a state ther- malised at one temperature to a new state eventually thermalised at a higher temperature. This is of interest to the lattice community who use the HMC algorithm for generating dynamical gauge ﬁeld ensembles, as it gives insight into the nature of the algorithm and correlation times.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	In generating the conﬁgurations used in this paper, we used 150 molecular dynamics steps with step size ∆τ = 1 150 . When producing visualisations we store the state of the Polyakov loops ﬁve times per trajectory in order to make the animation smoother. Of course, these frames are dropped if the conﬁguration is not accepted in step 3. The acceptance rates we observed when generating our gauge ﬁeld conﬁgurations varied between 75% and 90%, depending on the volume and temperature of the conﬁgurations being generated. Large volumes and high temperatures have the lowest acceptance rates.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	This ansatz for the static quark potential is sensitive to discretisation eﬀects at extremely small r and noise starts to dominate at large r, so we need a lower and upper cutoﬀ for r.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	On an anisotropic lattice, the ﬁt for the space-space loops (Wxy(r, t), Wxz(r, t), and Wyz(r, t)) gives a string tension σss which is related to the spatial lattice spacing as through the physical value √σ = 0.44 GeV. On the other hand, the space-time loops, (Wxt(r, t), Wyt(r, t), and Wzt(r, t)) and the time-space loops (Wtx(r, t), Wty(r, t), and Wtz(r, t)) give σst, which is related to both the spatial and temporal lattice spacings via √asat.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	choose a set of (β, γG) pairs that give us access to a range of temperatures at a ﬁxed volume as shown in Table 1.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	The universal properties at ﬁnite temperature phase transitions in (3 + 1) dimensional gauge theories are related to those in 3-dimensional spin models [16, 22]. Thus, it is of interest to compare the behavior of the Polyakov loops in QCD with the three dimensional 3-state Potts model [23], a generalisation of the Ising spin model. In the 3-state Potts model, each lattice site can assume one of three spin directions and these form spin aligned domains. After the phase transition, one direction dominates the space, just as in the QCD deconﬁnement transition. Thus, the Potts model is a candidate for a simpliﬁed model of deconﬁnement.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	is summed over the points of the lattice and the three spatial directions, ˆı is the unit vector in the positive i direction, and δj,k is the Kronecker delta. That is, E[σ] is the total number of spin-aligned nearest-neighbor pairs. Here, β = J/kBT is the inverse temperature in natural units.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	We can simulate this using the Metropolis-Hastings algorithm [24] to study the behavior of spin-aligned domains near the critical temperature and compare it to the behavior of center clusters in QCD near TC.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	In order to analyse the behavior of the Polyakov loops at the critical tem- perature, we generate gauge ﬁeld conﬁgurations on several 243 8 lattices with the same spatial lattice spacing but diﬀerent temporal spacings, using the pa- rameters described in Section 2.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Visualising the Polyakov loop data presents a diﬃculty. Most visualisation software that has the ability to render volumetric data takes in a three dimen- sional grid of real values and uses trilinear interpolation to ﬁll in the gaps. However, if we try to do this with the complex phase of our Polyakov loop val- ues (with a branch cut at π), we ﬁnd that the interpolation has problems near the branch cut. For example, if two adjacent data points have complex values 3 ) then interpolating − the phase linearly between the two points takes it through φ = 0 as in Fig. 2a, rather than crossing the branch cut as in Fig. 2b. This behavior is incorrect and produces artifacts in the ﬁnal image that look like thin shells or ﬁlms be- tween regions of transparency. This leads to much more red (corresponding to φ = 0) than either of green or blue (φ = 2π respectively) as seen in Fig. 3a.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Instead, we want to directly interpolate the complex numbers and then calcu- late the phase, resulting in the phase going directly between the two endpoints, across the branch cut as in Fig. 2b. This results in a signiﬁcantly diﬀerent visualisation as seen in Fig. 3b.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	To circumvent this problem, we have developed a custom volume renderer for complex valued scalar ﬁelds. The details of the rendering algorithm are given in Appendix A. There we present two ways of visualising the center clusters, one based on proximity of φ((cid:126)x) to one of the three center phases and one on the magnitude ρ((cid:126)x).
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Figure 3: Visualisations demonstrating the (red) artifacts produced by the incorrect phase interpolation on the left. On the right, the correct complex interpolation shows an equal center phase distribution.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Four animations are available online at the URL provided in Ref. [25]. This is the ﬁrst presentation of animations of the evolution of center clusters under the HMC algorithm. Two diﬀerent rendering techniques are presented. As discussed in Appendix A, these include rendering based on the proximity of φ((cid:126)x) to one of the three center phases of SU(3) and rendering based on the magnitude ρ((cid:126)x). Each are rendered from the original conﬁgurations as well as conﬁgurations smoothed with four sweeps of stout-link smearing as discussed in greater detail below.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	The animations reveal the evolution of center clusters as a function of the HMC simulation time, with ﬁve frames per unit trajectory. The evolution of the center clusters is governed by the evolution of the gauge ﬁeld conﬁgurations under the HMC algorithm, and therefore the timescale of these animations is not a physical scale but an algorithmic scale. However, while it may have no direct physical signiﬁcance, the evolution of the center clusters during thermalisation gives novel insights into the nature of the HMC algorithm and the way it brings the gauge ﬁeld to a physical conﬁguration. In addition, once thermalisation is complete, every frame in itself is a physically representative state and thus we can observe a number of possible structures for the center clusters and gain extensive insight.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Figure 4: Histogram of the phase φ((cid:126)x) and a visualisation of the center clusters in the local Polyakov loop values at T = 0.89(1) TC . All three center phases are present in small clusters with approximately equal density.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Figure 5: Histogram and visualisation as in Fig. 4 at T = 1.14(2) TC . A single(red) phase is beginning to dominate, signaling deconﬁnement.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	hand plot, a histogram of the distribution of the phases of the Polyakov loops across a gauge ﬁeld ensemble at this temperature shows that all three center phases, observable as peaks in the histogram, are approximately equally occu- pied, signaling conﬁnement. The right-hand plot shows a single frame from the animation, in which the three center-phase peaks observed in the histogram cor- respond to blue (left), red (center), and green (right). All three center phases are present in small clusters with approximately equal density.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	At this point (0:30 in the animation), the temperature is increased to T = 307(4) MeV or T /TC = 1.14(2) and the response of the gauge ﬁeld is illustrated in the animations. Fig. 5 shows the red center phase becoming dominant with the other two beginning to be suppressed, signaling the onset of deconﬁnement. In the animation, we can see that the three center phases start out equally dominant and ﬂuctuate in size until the red clusters come to dominate.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Figure 6: Histogram and visualisation at T = 1.36(2) TC . A single (red) cluster dominates the entire space.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Figure 7: Histogram and visualisation at T = 1.61(3) TC . The same (red) phase is still dominant.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	temperature again, this time to T = 366(5) MeV or T /TC = 1.36(2). At this temperature, the animation shows that the red clusters continue to grow, oc- cupying almost the entire space. This can be seen in the snapshot provided in Fig. 6.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	After 250 more HMC trajectories or 50 seconds, at 3:20 in the animation, we increase the temperature again, to T = 434(7) MeV or T /TC = 1.61(3). At this temperature, the red phase remains dominant in the animation and the other two phases are suppressed even further. Fig. 7 provides a snapshot at this temperature, showing the red phase almost completely dominant. We show 250 HMC trajectories at this temperature over the remaining 50 seconds of the animation.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Figure 8: Evolution of center clusters with simulation time at T = 0.89(1) TC .
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	ent observables under HMC evolution can vary, so it is of interest to observe the timescales over which the center clusters evolve. We can see that over the course of a single unit trajectory, the small scale structure of the loops changes signiﬁcantly, as shown in Fig. 8. This suggests that the small scale structure of the center clusters evolves very quickly under the HMC algorithm and thus has short autocorrelation times.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	To investigate the larger scale behavior of the clusters, we remove the small scale noise by performing four sweeps of stout-link smearing [26] prior to cal- culating the Polyakov loops. The results are illustrated in Fig. 9. In these visualisations, we see that over the course of one trajectory, the center clusters are slowly moving, with some change around the boundaries of the clusters. Observing the evolution of the center clusters in the corresponding animation, we see correlations in the center clusters persisting for approximately 5 sec- onds corresponding to 25 trajectories, suggesting an approximate length for the autocorrelation time of the larger scale structure of the clusters. After ap- proximately 50 trajectories, the large scale structure of the loops has become completely decorrelated. This is supported by Fig. 10, which shows the phase of the average Polyakov loop on a small (6 6) subsection of the lattice. We can see that the phase becomes completely decorrelated within 50 trajectories. This supports our choice of separation for independent conﬁgurations.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	The evolution observed in SU (3) gauge theory is consistent with the be- havior of spin-aligned domains in the three dimensional 3-state Potts model [23] just below the critical temperature under Metropolis-Hastings algorithm simulations, as seen in Fig. 11.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Figure 9: Evolution of center clusters with simulation time at T = 0.89(1) TC . Four sweeps of stout-link smearing are applied to the gauge links prior to calculating the Polyakov loops.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Figure 11: Evolution of spin-aligned domains in the three dimensional 3-state Potts model at β = 0.55 on a 163 lattice.
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	If we use the alternate rendering style based on the magnitude ρ(x), we can see that peaks in the magnitude lie approximately within the center clusters, as shown in Fig. 12. The magnitude peaks are each colored corresponding to a single center phase and they appear to line up with peaks in the corresponding phase based visualisation. However it is not necessarily clear that each center cluster corresponds to a peak in the modulus, ρ((cid:126)x).
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	In order to look more closely at the correlation between these clusters, we again perform four sweeps of stout-link smearing, making it easier to observe larger scale structures, and render both the phase and magnitude based clus- ters using an isovolume rendering developed in AVS/Express [27] that makes the boundaries of the clusters more clear. When we do this, it becomes clear that there is an approximate one-to-one relationship between the peaks in the magnitude ρ((cid:126)x) and the proximity of φ((cid:126)x) to a center phase. The measures /(π/3) and ρ(x)/ρmax are illustrated in Figs. 13 and 14. We φc((cid:126)x) φ((cid:126)x) \| \| can also observe this correlation by looking at a scatter plot of ρ(x)/ρmax vs as in Fig. 15. This observed correlation of ρ((cid:126)x) becoming small φ((cid:126)x) \| as φ((cid:126)x) moves away from a center phase is the ﬁrst direct conﬁrmation of the underlying assumption of Ref. [5], which links the center domain walls to unan- ticipated phenomena observed at RHIC [28] and the LHC [6–8].
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Figure 12: Comparison of complex phase φ((cid:126)x) and magnitude ρ((cid:126)x) clusters at T = 0.89(1) TC .
Processed_Visualisations_of_coherent_centre_domains_in_local.txt	Figure 13: Comparison of phase φ((cid:126)x) and magnitude ρ((cid:126)x) clusters at T = 0.89(1) TC after four sweeps of stout-link smearing using the isovolume renderer. The rendering thresholds for φ((cid:126)x) and ρ((cid:126)x) are 0.5 and 0.2 respectively.

Subsets and Splits