Search

We consider the problem of neural network training in a time-varying context. Machine learning algorithms have excelled in problems that do not change over time. However, problems encountered in financial markets are often time varying. We propose the online early stopping algorithm and show that a neural network trained using this algorithm can track a function changing with unknown dynamics. We compare the proposed algorithm to current approaches on predicting monthly US stock returns and show its superiority. We also show that prominent factors (such as the size and momentum effects) and industry indicators exhibit time-varying predictive power on stock returns. We find that during market distress, industry indicators experience an increase in importance at the expense of firm level features. This indicates that industries play a role in explaining stock returns during periods of heightened risk.

Financial markets are noisy learning environments. We propose an approach that regularizes the Temporal Convolutional Network using a supervised autoencoder, which we term the Supervised Temporal Autoencoder (STAE). We show that the addition of the auxiliary reconstruction task is beneficial to the primary supervised learning task in the context of stock return time-series forecasting. We also show that STAE is able to learn features directly from transformed price series, alleviating the need for handcrafted features. The autoencoder also improves interpretability as users can observe output of the decoder and inspect features retained by the network.

This paper proposes a new method for determining similarity and anomalies between time series, most practically effective in large collections of (likely related) time series, by measuring distances between structural breaks within such a collection. We introduce a class of semi-metric distance measures, which we term MJ distances. These semi-metrics provide an advantage over existing options such as the Hausdorff and Wasserstein metrics. We prove they have desirable properties, including better sensitivity to outliers, while experiments on simulated data demonstrate that they uncover similarity within collections of time series more effectively. Semi-metrics carry a potential disadvantage: without the triangle inequality, they may not satisfy a "transitivity property of closeness." We analyse this failure with proof and introduce an computational method to investigate, in which we demonstrate that our semi-metrics violate transitivity infrequently and mildly. Finally, we apply our methods to cryptocurrency and measles data, introducing a judicious application of eigenvalue analysis.

© 2018 Bell et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Objective Blood pressure variability (BPV) has been associated with risk of cardiovascular events in observational studies, independently of mean BP levels. In states with higher autonomic imbalance, such as in diabetes, the importance of BP variability may theoretically be even greater. We aimed to investigate the incremental value of BPV for prediction of cardiovascular and all-cause mortality in patients with type 2 diabetes. Methods We identified 9,855 patients without pre-existing cardiovascular disease who did not change BP-lowering treatment during the observation period from a Swedish primary health care cohort of patients with type 2 diabetes. BPV was summarized as the standard deviation (SD), coefficient of variation (CV), or variation independent of mean (VIM). Patients were followed for a median of 4 years and associations with cardiovascular and all-cause mortality were investigated using Cox proportional hazards models. Results BPV was not associated with cardiovascular specific or all-cause mortality in the total sample. In patients who were not on BP-lowering drugs during the observation period (n = 2,949), variability measures were associated with all-cause mortality: hazard ratios were 1.05, 1.04 and 1.05 for 50% increases in SD, CV and VIM, respectively, adjusted for Framingham risk score risk factors, including mean BP. However, the addition of the variability measures in this subgroup only led to very minimal improvement in discrimination, indicating they may have limited clinical usefulness (change in C-statistic ranged from 0.000–0.003 in all models). Conclusions Although BPV was independently associated with all-cause mortality in diabetes patients in primary care who did not have pre-existin

Item does not contain fulltext

Item does not contain fulltext, During speech listening, the brain parses a continuous acoustic stream of information into computational units (e.g. syllables or words) necessary for speech comprehension. Recent neuroscientific hypotheses propose that neural oscillations contribute to speech parsing, but whether they do so on the basis of acoustic cues (bottom-up acoustic parsing) or as a function of available linguistic representations (top-down linguistic parsing) is unknown. In this magnetoencephalography study, we contrasted acoustic and linguistic parsing using bistable speech sequences. While listening to the speech sequences, participants were asked to maintain one of the two possible speech percepts through volitional control. We predicted that the tracking of speech dynamics by neural oscillations would not only follow the acoustic properties but also shift in time according to the participant's conscious speech percept. Our results show that the latency of high-frequency activity (specifically, beta and gamma bands) varied as a function of the perceptual report. In contrast, the phase of low-frequency oscillations was not strongly affected by top-down control. While changes in low-frequency neural oscillations were compatible with the encoding of pre-lexical segmentation cues, high-frequency activity specifically informed on an individual's conscious speech percept.

Diagnosis of the speech motor planning/programming disorder, apraxia of speech (AOS), has proven challenging, largely due to its common co-occurrence with the language-based impairment of aphasia. Currently, diagnosis is based on perceptually identifying and rating the severity of several speech features. It is not known whether all, or a subset of the features, are required for a positive diagnosis. The purpose of this study was to assess predictor variables for the presence of AOS after left-hemisphere stroke, with the goal of increasing diagnostic objectivity and efficiency. This population-based case-control study involved a sample of 72 cases, using the outcome measure of expert judgment on presence of AOS and including a large number of independently collected candidate predictors representing behavioral measures of linguistic, cognitive, nonspeech oral motor, and speech motor ability. We constructed a predictive model using multiple imputation to deal with missing data; the Least Absolute Shrinkage and Selection Operator (Lasso) technique for variable selection to define the most relevant predictors, and bootstrapping to check the model stability and quantify the optimism of the developed model. Two measures were sufficient to distinguish between participants with AOS plus aphasia and those with aphasia alone, (1) a measure of speech errors with words of increasing length and (2) a measure of relative vowel duration in three-syllable words with weak-strong stress pattern (e.g., banana, potato). The model has high discriminative ability to distinguish between cases with and without AOS (c-index=0.93) and good agreement between observed and predicted probabilities (calibration slope=0.94). Some caution is warranted, given the relatively small sample specific to left-hemisphere stroke, and the limitations of imputing missing data. These two speech measures are straightforward to collect and analyse, facilitating use in research and clinical settings.