Echo State Networks for Bitcoin Time Series Prediction

Enrico Sartor; Helmut Prendinger; Mansi Sharma; Marc Cavazza

REVIEW 5 major objections 5 minor 30 references

Echo State Networks for Bitcoin Time Series Prediction

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Echo state networks with a fixed reservoir and a trained readout beat XGBoost and the naive baseline on one-day-ahead Bitcoin closing prices, and the gap widens in windows the maximal Lyapunov exponent marks as highly chaotic.

desk verdict Useful empirical comparison of ESN vs XGBoost on Bitcoin with a chaos-conditional twist, but the headline claims outrun the evidence and Section 4 contains a large verbatim block from another paper that needs urgent explanation. read the letter →

arxiv 2508.05416 v1 pith:5OZD2UP4 submitted 2025-08-07 cs.LG cs.CEcs.NE

Mansi Sharma , Enrico Sartor , Marc Cavazza , Helmut Prendinger This is my paper

classification cs.LGcs.CEcs.NE

keywords echostatenetworksreservoircomputingBitcoinpricepredictioncryptocurrencyforecastingLyapunovexponentchaotictimeseriesXGBoost

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that a tuned echo state network—a recurrent model with a fixed random reservoir and only its output layer trained—forecasts next-day Bitcoin closing prices more accurately than XGBoost or the naive last-value rule on daily data from August 2017 to January 2023. The comparison is run on rolling training windows of 15, 30, and 60 days, in univariate and multivariate settings, and the ESN has the lowest mean RMSE in every configuration. The paper's second claim is that the advantage is not uniform: in windows where a small-data Lyapunov-exponent estimate indicates strong chaos, the ESN beats XGBoost by a statistically significant margin (a non-parametric rank-sum test at the 5% level, for 30- and 60-day windows). If right, this matters because a cheap chaos measure could say when reservoir computing will help most, and because it undercuts the idea that highly volatile crypto prices are too chaotic to predict at all. The authors note that the 15-day windows do not reproduce the effect, likely because Lyapunov estimates from so few points are unreliable.

What carries the argument

The central object is the echo state network itself: a reservoir of randomly connected, fixed neurons whose state evolves as $\tilde{x}(k)=\tanh(W_{\mathrm{in}}u(k)+W_x x(k-1))$, with the prediction $y(k)=W_{\mathrm{out}}(\alpha\tilde{x}(k)+(1-\alpha)x(k-1))$ and only $W_{\mathrm{out}}$ learned by linear regression. The chaos-conditioned claim is carried by the maximal Lyapunov exponent $\lambda_{\max}$ computed from each rolling window using the small-data algorithm of the cited work, then averaged over the ten train-test splits in that window. The statistical verdict rests on the ratio $\mathrm{EXratio}=\mathrm{RMSE}_{\mathrm{ESN}}/\mathrm{RMSE}_{\mathrm{XGB}}$, whose high-chaos versus low

What would settle it

Compute the maximal Lyapunov exponent for the same windows with several embedding dimensions and time delays, and re-run the rank-sum test on the resulting high-/low-chaos split; also run the split using realized volatility instead of MLE. If the significant ESN advantage moves or disappears under these perturbations, the chaos-conditional claim is an artifact of the estimator rather than a property of the market.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a tuned ESN gives the lowest mean RMSE in every configuration tested: for 15-, 30-, and 60-day training windows, ESN mean RMSE is 442.78, 463.96, and 427.35 dollars univariate and 487.77, 533.43, and 441.23 dollars multivariate, versus XGBoost's 740.43, 741.84, 675.95 and 678.59, 676.60, 572.58, with the naive baseline higher still. ESN errors are 1.0–2.7% of closing price, averaging 1.8%. The chaos-conditional claim is that $\mathrm{EXratio}=\mathrm{RMSE}_{\mathrm{ESN}}/\mathrm{RMSE}_{\mathrm{XGB}}$ has a significantly lower median in high-chaos than in low-chaos windows for sizes 30 and 60, meaning the ESN is relatively strongest where the Lyapu

Load-bearing premise

The load-bearing premise is that the maximal Lyapunov exponent computed by the small-data algorithm over short rolling windows of 15-60 daily Bitcoin prices genuinely measures how chaotic the market dynamics are; if those estimates are unstable or the price series is not a low-dimensional deterministic system, the claim that ESNs excel under high chaos has no foundation.

Editorial extensions

If this is right

On the paper's data, an ESN predicts next-day Bitcoin closes with mean RMSE between about 427 and 533 USD across window sizes and feature sets, beating XGBoost and the naive baseline in every configuration studied.
The ESN's advantage over XGBoost is statistically significant in the windows with the highest Lyapunov exponents (30- and 60-day training windows), so the model is relatively strongest during chaotic periods, not weakest.
Multivariate inputs—moving averages, RSI, MACD, price-range and gap features, and trend coefficients—generally lower RMSE for both ESN and XGBoost compared with closing prices alone.
The 15-day training windows do not show the chaos-conditional effect, so the paper's practical implication is that chaos measures from fewer than about 30 daily points should not be trusted for model selection.
At the one-day horizon, the results are evidence against the strong efficient-markets hypothesis: a lightweight recurrent model extracts predictable structure from a volatile, non-stationary crypto series.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension the authors do not pursue: use $\lambda_{\max}$ measured on a rolling window as a switching signal between ESN and XGBoost, so the more accurate model is chosen per regime.
The 15-window failure is consistent with an estimation artifact, but it could also mean the reservoir advantage is horizon-dependent; testing the same protocol on longer histories or intraday data would separate those readings.
Because only XGBoost and the naive rule enter the statistical test, it is untested whether LSTM or transformer models also excel under high chaos; if they do, the claim is about flexible nonlinear models rather than reservoir computing specifically.
The chaos split itself is defined by an estimator of deterministic dynamics on a series with jumps and regime shifts; replacing MLE with realized volatility or sample entropy would test whether chaos, rather than volatility, is the operative variable.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the paper does something useful—rolling-window ESN versus XGBoost and naive baselines on daily BTCUSDT data, plus a Mann-Whitney test linking ESN's relative advantage to windows with high maximal Lyapunov exponent. The empirical direction is defensible: mean RMSE for ESN is lower than the baselines across the reported windows, and the chaos-conditional analysis is a genuine addition. The problem is that the paper's claims are bigger than the evidence supports.

What's actually new: the specific chaos-conditional comparison. Kenny (2020) already applied ESNs to Bitcoin and Ethereum, so the 'first exploration' framing is overstated, and the authors admit this in related work. The added value is the systematic tuning and the Lyapunov-conditioned test, which is a legitimate extension. Credit where due: the Mann-Whitney design is sound, and the limitation section honestly notes the 15-day window instability.

Soft spots, in proportion:

First, the headline margin has no statistical support. Table 3 shows lower mean RMSE, but no error bars and no significance test; the word 'significant' is doing too much work. Fixable.

Second, the evaluation protocol is under-specified. The paper says data is pre-split into windows with training sizes 15/30/60 and test size 10, but does not state that hyperparameter tuning (TPE) is nested within the training part of each window. If tuning or feature selection touched the test folds, the RMSEs are optimistically biased. This is more concerning than the MLE reliability issue.

Third, the multivariate claim contradicts the paper's own table. The text says multivariate models 'generally perform better' and 'particularly true for the ESN,' but Table 3 shows ESN univariate RMSEs of 442/464/427 versus multivariate 488/533/441—worse in every window.

Fourth, and most seriously, Section 4.1 contains a long verbatim block from Viehweg, Worthmann, and Mäder (Neurocomputing 2023), including their introduction and equations. It reads as an accidental LaTeX inclusion, but as published it is unattributed copying and makes the methods section unusable. This must be addressed before anything else.

The MLE analysis is fragile on short windows—the authors admit the 15-day case is not significant—but that is a minor caveat since the effect appears at 30 and 60 days. No code is provided, which also limits verification.

Who this is for: reservoir computing people and anyone working on crypto forecasting. It deserves peer review because there is a real empirical question here, but it needs major revision: clean the copied text, clarify the tuning procedure, add uncertainty quantification, and align the claims with the evidence.

I would not cite it in its current form. My recommendation: send to review, but expect heavy revision.

Referee Report

5 major / 5 minor

Summary. The paper applies Echo State Networks (ESNs) to one-day-ahead forecasting of daily Bitcoin closing prices, comparing against XGBoost and a Naïve Last-Value baseline. The evaluation uses a rolling-window setup with training sizes of 15, 30, and 60 points, a fixed test size of 10 points, and stride 1, in both univariate (closing price) and multivariate (with hand-selected technical features) settings. Hyperparameters for ESN and XGBoost are tuned with Tree Parzen Estimators. The authors report that ESN achieves the lowest mean RMSE in every setting (Table 3) and claim, based on a Mann-Whitney test of the ESN/XGBoost RMSE ratio against the Lyapunov exponent, that ESN performs significantly better in high-chaos windows.

Significance. If the empirical protocol is sound, the paper would provide one of the first extensive evaluations of ESNs for cryptocurrency forecasting and a potentially useful chaos-conditional analysis. The use of public Binance data, explicit rolling-window splits, and a reproducible TPE search space are assets. However, the headline claims hinge on the integrity of the evaluation protocol and on the reliability of the Lyapunov-exponent-based chaos measure; both currently rest on ambiguous or under-specified methodology.

major comments (5)

[§5, Hyperparameter Selection; §4.3, Features] The evaluation protocol is ambiguous with respect to data leakage. Section 5 states that TPE searches for hyperparameter combinations, but it does not state that this search is nested inside the training portion of each rolling window. Likewise, §4.3 says features were 'evaluated for their impact on prediction accuracy across time windows,' which could mean the held-out test windows were used for feature selection. If hyperparameter optimization or feature selection uses any test-fold information, the RMSE values in Table 3 are optimistically biased and the reported ESN margin over XGBoost and Naïve may be an artifact. The manuscript must specify, for each fold, exactly which data are used for training, validation, and testing, and should ideally provide code or pseudo-code. This is load-bearing for every quantitative claim in the paper.
[§6.1, Chaos analysis] The Mann-Whitney U test is applied to EXratio values grouped by high/low chaos, but the windows are constructed with stride 1 and are therefore heavily overlapping. The test assumes independent observations, which is violated: consecutive windows share almost all training and test points, and the Lyapunov exponents are computed on overlapping data (the text even says the exponent is computed 10 times per window and averaged). The reported p-values are therefore not valid evidence for the claim that ESN outperforms XGBoost more in high-chaos windows. The authors should use non-overlapping or block-bootstrapped windows, or explicitly justify independence.
[§4.1, ESN formulation] Section 4.1 contains a large, verbatim passage on ESN hyperparameter tuning from Viehweg et al. (2023, Neurocomputing), which is repeated multiple times and is not integrated with the paper's own contribution. This appears to be a copy-paste error and is a serious presentation/integrity problem. It must be removed and replaced with an original description of the ESN equations and tuning setup. The duplicated text also interrupts the flow between Eq. (1) and Eq. (2).
[Abstract; §6, Results] The abstract claims ESNs 'outperform existing machine learning methods by a significant margin,' but the supporting evidence is only the mean RMSE values in Table 3, with no error bars, confidence intervals, or significance test on the RMSE comparisons across windows. Given the large variance visible in Figures 4–6, a paired bootstrap or a signed-rank test across folds is needed to support the phrase 'significant margin.'
[§6.1, Lyapunov exponent computation] The chaos-conditional claim rests on the Maximal Lyapunov Exponent computed by Rosenstein's algorithm on rolling windows of only 15–60 daily price points. The authors themselves note the failure for window size 15 and attribute it to the limited number of points. For windows of 30 and 60 points, the MLE estimate is still based on very short trajectories and may reflect noise rather than deterministic chaos. The authors should validate the MLE estimates with surrogate-data tests or report robustness to embedding parameters; otherwise the 'excels under high chaos' conclusion is not solidly grounded.

minor comments (5)

[§4.1, Eq. (1)] Typos: 'Jager' should be 'Jaeger'; the reference list also contains 'Lukoeviius' for 'Lukoševičius'.
[Table 3] Table 3 is not formatted as a numbered table in the text; it appears as a raw text block. The caption and column structure should be fixed, and the table should be referenced consistently.
[§5, Evaluation] The phrase 'cross-validation inside each window' is unclear: the setup described is a rolling-window forecast, not k-fold cross-validation. Please use consistent terminology and clarify whether the test set is always the 10 points immediately following the training window.
[§3, Dataset] The paper states 'we considered standardization and converting prices to returns' but does not report the results of these preliminary experiments. Either report them in a supplementary table or omit the mention.
[§4.3, Features] The feature list is descriptive but not accompanied by a summary table with names and formulas. Adding a table would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ESN results are benchmarked against external baselines on public data, and the chaos-conditional analysis is an empirical association, not a definitional reduction.

full rationale

The paper's central claims rest on out-of-sample prediction error (RMSE) computed on fixed 10-day test windows within a rolling-window protocol, comparing ESN against XGBoost and Naïve baselines on the same data. No equation in the paper defines the predicted quantity in terms of the fitted parameters or the Lyapunov exponent; the ESN output weights are trained by linear regression on training windows and evaluated on separate test points. Hyperparameter tuning via TPE is applied symmetrically to ESN and XGBoost and is a standard model-selection step, not a fitted-value-reused-as-prediction. The chaos analysis computes the Maximal Lyapunov exponent from the price series independently of model errors, then tests whether the ESN/XGBoost RMSE ratio differs between high- and low-chaos windows; this is a conditional performance comparison, not a derivation of performance from the chaos measure itself. There are no self-citations that carry load-bearing uniqueness claims, no imported ansatz masquerading as first principles, and no renaming of a known result as a new derivation. The acknowledged limitations (short windows; difficulty computing Lyapunov exponents for 15-day windows) and the potential ambiguity about whether feature selection was nested within each rolling window are legitimate methodological risks, but they do not constitute circularity under the specified criteria because no claimed result is equivalent to its own input by construction.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central comparison depends on a large set of tuned hyperparameters and on the interpretation of MLE over very short windows. No new entities are introduced. The main unstated cost is that hyperparameter and feature selection happened on the same data used for evaluation.

free parameters (4)

ESN hyperparameters (reservoir size, spectral radius, leaking rate, output weight regularization, reservoir/input/feedba = not reported (search ranges in Table 1)
Tuned on the Bitcoin data via Tree Parzen Estimators; exact selected values are not given.
XGBoost hyperparameters (n_estimators, max_depth, gamma, learning_rate, colsample_by_tree, min_child_weight, subsample) = not reported (search ranges in Table 2)
Tuned by the same TPE procedure on the same data; exact values are absent.
Lyapunov exponent embedding parameters (embedding dimension, time delay, trajectory length) = not reported
Rosenstein's algorithm requires these; the paper reports only the algorithm and window sizes, so the MLE values are not uniquely reproducible.
Training window sizes and test size = 15, 30, 60 days training; 10 days test
Discrete design choices that materially affect the result; the 15-day case fails the chaos test.

assumptions (3)

domain assumption The ESN reservoir satisfies the echo state property after an initial transient, so output weights can be fit by linear regression (Eq. 3).
Standard reservoir computing assumption; not verified for these random matrices and inputs.
domain assumption Daily Bitcoin close prices can be treated as a realization of a low-dimensional deterministic chaotic system for the purpose of Lyapunov estimation.
Rosenstein MLE assumes an underlying attractor; financial series are stochastic, so MLE values are interpretively fragile.
ad hoc to paper Rolling windows used in the Mann-Whitney test are independent samples.
Windows slide with stride 1 and a 10-point test set, so consecutive windows share data; the independence assumption of the statistical test is violated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Echo State Networks for Bitcoin Time Series Prediction." pith.science (2026). https://pith.science/paper/5OZD2UP4

@misc{pith2026250805416,
  author       = {Pith},
  title        = {Pith review of: Echo State Networks for Bitcoin Time Series Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5OZD2UP4}},
  note         = {Machine review of arXiv:2508.05416}
}

read the original abstract

Forecasting stock and cryptocurrency prices is challenging due to high volatility and non-stationarity, influenced by factors like economic changes and market sentiment. Previous research shows that Echo State Networks (ESNs) can effectively model short-term stock market movements, capturing nonlinear patterns in dynamic data. To the best of our knowledge, this work is among the first to explore ESNs for cryptocurrency forecasting, especially during extreme volatility. We also conduct chaos analysis through the Lyapunov exponent in chaotic periods and show that our approach outperforms existing machine learning methods by a significant margin. Our findings are consistent with the Lyapunov exponent analysis, showing that ESNs are robust during chaotic periods and excel under high chaos compared to Boosting and Na\"ive methods.

Figures

Figures reproduced from arXiv: 2508.05416 by the authors.

**Figure 2.** Bitcoin differenced prices. to apply ESNs to cryptocurrency forecasting but without extensive ESN tuning, which is the focus of our work. 3 Dataset and Descriptive Statistics We used daily Bitcoin data evaluated in US dol lars (ticker: BTCUSDT) sourced from the Binance trading platform [4]. Missing values were filled using linear interpolation, with no additional transformations applied, though we considered standar… view at source ↗

**Figure 1.** Building blocks of an ESN: layered structure, core parameters, and their dimensionality. J. Viehweg, K. Worthmann and P. Mäder Neurocomputing 522 (2023) 214–228 215 of randomly generated weight parameters into account. We focus on the ESN as proposed by Jaeger and his group [48,50,62] including the leakage concept, a promising ESN evolution. We apply the proposed optimization procedure on the challenging task of pr… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: RMSE values on uni- and multi-variate setting using a [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: RMSE values on uni- and multi-variate setting using a [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 6.** Figure 6: RMSE values on uni- and multi-variate setting using a [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 7.** Figure 7: (Top) Amount of chaos measured by Lyapunov exponent, [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

**Figure 8.** Figure 8: Chaos analysis 7 Discussion We showed that ESN is particularly effective in forecasting Bitcoin prices during highly chaotic market conditions, as highlighted by the Lyapunov [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 29 canonical work pages

[1]

Complexity 2018, 1–16 (nov 2018)

Alessandretti, L., ElBahrawy, A., Aiello, L.M., Baronchelli, A.: An- ticipating cryptocurrency prices using machine learning. Complexity 2018, 1–16 (nov 2018)

work page 2018
[2]

Expert systems with applications140, 112896 (2020)

Bandara, K., Bergmeir, C., Smyl, S.: Forecasting across time series databases using recurrent neural networks on groups of similar se- ries: A clustering approach. Expert systems with applications140, 112896 (2020)

work page 2020
[3]

Economics Letters161, 1–4 (2017)

Bariviera, A.F.: The inefficiency of bitcoin revisited: A dynamic ap- proach. Economics Letters161, 1–4 (2017)

work page 2017
[4]

Binance: Binance: The world’s leading cryptocurrency exchange (2025)

work page 2025
[5]

Applied Soft Computing90, 106187 (2020)

Borges, T.A., Neves, R.F.: Ensemble of machine learning algorithms for cryptocurrency investment with different data resampling meth- ods. Applied Soft Computing90, 106187 (2020)

work page 2020
[6]

Energy165, 709–726 (2018)

Chou, J.S., Tran, D.S.: Forecasting energy consumption time series using machine learning techniques based on usage patterns of resi- dential householders. Energy165, 709–726 (2018)

work page 2018
[7]

Abstract and Applied Analysis 2014, 1–6 (06 2014)

Dan,J.,Guo,W.,Shi,W.,Fang,B.,Tingping,Z.:Deterministicecho state networks based stock price forecasting. Abstract and Applied Analysis 2014, 1–6 (06 2014)

work page 2014
[8]

Elsayed,S.,Thyssens,D.,Rashed,A.,Jomaa,H.S.,Schmidt-Thieme, L.: Do we really need deep learning models for time series forecast- ing? arXiv preprint arXiv:2101.02118 (2021)

arXiv 2021

Show all 30 references

[9]

European Journal of Operational Research270(2), 654–669 (2018)

Fischer, T., Krauss, C.: Deep learning with long short-term mem- ory networks for financial market predictions. European Journal of Operational Research270(2), 654–669 (2018)

2018
[10]

Science304, 78– 80 (2004)

Jaeger, H., Haas, H.: Harnessing nonlinearity: Predicting chaotic sys- tems and saving energy in wireless communication. Science304, 78– 80 (2004)

2004
[11]

In: Wirtschaftsinformatik (Zentrale Tracks)

Jaquart, P., Dann, D., Martin, C.: Machine learning for bitcoin pricing-a structured literature review. In: Wirtschaftsinformatik (Zentrale Tracks). pp. 174–188 (2020)

2020
[12]

The Journal of Finance and Data Science (2021)

Jaquart, P., Dann, D., Weinhardt, C.: Short-term bitcoin market prediction via machine learning. The Journal of Finance and Data Science (2021)

2021
[13]

Master’s thesis, Technological University Dublin, Dublin, Ireland (2020) Echo State Networks for Bitcoin Time Series Prediction 15

Kenny, M.: Investigating the Predictability of a Chaotic Time-Series Data using Reservoir Computing, Deep-Learning and Machine- Learning on the Short-, Medium-and Long-Term Pricing of Bitcoin and Ethereum. Master’s thesis, Technological University Dublin, Dublin, Ireland (2020...

2020
[14]

Energies8(10), 12228–12241 (2015)

Li, G., Li, B.J., Yu, X.G., Cheng, C.T.: Echo state network with bayesian regularization for forecasting short-term power production of small hydropower plants. Energies8(10), 12228–12241 (2015)

2015
[15]

Expert System Applications36(3), 7313–7317 (apr 2009)

Lin, X., Yang, Z., Song, Y.: Short-term stock price prediction based on echo state networks. Expert System Applications36(3), 7313–7317 (apr 2009)

2009
[16]

In: Neural Networks (2012)

Lukoeviius, M.: A practical guide to applying echo state networks. In: Neural Networks (2012)

2012
[17]

PLOS ONE13(3), 1–26 (2018)

Makridakis, S., Spiliotis, E., Assimakopoulos, V.: Statistical and machine learning forecasting methods: Concerns and ways forward. PLOS ONE13(3), 1–26 (2018)

2018
[18]

Jour- nal of the Operational Research Society74(3), 840–859 (2023)

Makridakis, S., Spiliotis, E., Assimakopoulos, V., Semenoglou, A.A., Mulder, G., Nikolopoulos, K.: Statistical, machine learning and deep learning forecasting methods: Comparisons and ways forward. Jour- nal of the Operational Research Society74(3), 840–859 (2023)

2023
[19]

Economics Letters 150, 6–9 (2017)

Nadarajah, S., Chu, J.: On the inefficiency of bitcoin. Economics Letters 150, 6–9 (2017)

2017
[20]

Bitcoin.–URL: https://bitcoin

Nakamoto, S., Bitcoin, A.: A peer-to-peer electronic cash system. Bitcoin.–URL: https://bitcoin. org/bitcoin. pdf4(2), 15 (2008)

2008
[21]

Wiley Finance, Wiley (1994)

Peters, E.: Fractal Market Analysis: Applying Chaos Theory to In- vestment and Economics. Wiley Finance, Wiley (1994)

1994
[22]

dePrado,M.:AdvancesinFinancialMachineLearning.Wiley(2018)

2018
[23]

Big Data Research23, 100178 (2021)

Ren, X., Li, X., Ren, K., Song, J., Xu, Z., Deng, K., Wang, X.: Deep learning-based weather prediction: a survey. Big Data Research23, 100178 (2021)

2021
[24]

Physica D: Nonlinear Phenomena65(1-2), 117–134 (1993)

Rosenstein,M.T.,Collins,J.J.,DeLuca,C.J.:Apracticalmethodfor calculating largest lyapunov exponents from small data sets. Physica D: Nonlinear Phenomena65(1-2), 117–134 (1993)

1993
[25]

International Journal of Forecasting36 (07 2019)

Smyl, S.: A hybrid method of exponential smoothing and recurrent neural networks for time series forecasting. International Journal of Forecasting36 (07 2019)

2019
[26]

International Journal of Forecasting35 (11 2018)

Spiliotis, E., Nikolopoulos, K., Assimakopoulos, V.: Tales from tails: On the empirical distributions of forecasting errors and their impli- cation to risk. International Journal of Forecasting35 (11 2018)

2018
[27]

Finance Research Letters32, 101084 (2020)

Sun, X., Liu, M., Sima, Z.: A novel cryptocurrency price trend forecasting model based on lightgbm. Finance Research Letters32, 101084 (2020)

2020
[28]

Economics Letters148(C), 80–82 (2016)

Urquhart, A.: The inefficiency of bitcoin. Economics Letters148(C), 80–82 (2016)

2016
[29]

The Journal of Financial Data Science4(1), 111–129 (dec 2021)

Wood, K., Roberts, S., Zohren, S.: Slow momentum with fast re- version: A trading strategy using deep learning and changepoint de- tection. The Journal of Financial Data Science4(1), 111–129 (dec 2021)

2021
[30]

Journal of Algorithms & Computational Technology7(1), 87–100 (2013)

Zhang, H., Liang, J., Chai, Z.: Stock prediction based on phase space reconstruction and echo state networks. Journal of Algorithms & Computational Technology7(1), 87–100 (2013)

2013

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Complexity 2018, 1–16 (nov 2018)

Alessandretti, L., ElBahrawy, A., Aiello, L.M., Baronchelli, A.: An- ticipating cryptocurrency prices using machine learning. Complexity 2018, 1–16 (nov 2018)

work page 2018

[2] [2]

Expert systems with applications140, 112896 (2020)

Bandara, K., Bergmeir, C., Smyl, S.: Forecasting across time series databases using recurrent neural networks on groups of similar se- ries: A clustering approach. Expert systems with applications140, 112896 (2020)

work page 2020

[3] [3]

Economics Letters161, 1–4 (2017)

Bariviera, A.F.: The inefficiency of bitcoin revisited: A dynamic ap- proach. Economics Letters161, 1–4 (2017)

work page 2017

[4] [4]

Binance: Binance: The world’s leading cryptocurrency exchange (2025)

work page 2025

[5] [5]

Applied Soft Computing90, 106187 (2020)

Borges, T.A., Neves, R.F.: Ensemble of machine learning algorithms for cryptocurrency investment with different data resampling meth- ods. Applied Soft Computing90, 106187 (2020)

work page 2020

[6] [6]

Energy165, 709–726 (2018)

Chou, J.S., Tran, D.S.: Forecasting energy consumption time series using machine learning techniques based on usage patterns of resi- dential householders. Energy165, 709–726 (2018)

work page 2018

[7] [7]

Abstract and Applied Analysis 2014, 1–6 (06 2014)

Dan,J.,Guo,W.,Shi,W.,Fang,B.,Tingping,Z.:Deterministicecho state networks based stock price forecasting. Abstract and Applied Analysis 2014, 1–6 (06 2014)

work page 2014

[8] [8]

Elsayed,S.,Thyssens,D.,Rashed,A.,Jomaa,H.S.,Schmidt-Thieme, L.: Do we really need deep learning models for time series forecast- ing? arXiv preprint arXiv:2101.02118 (2021)

arXiv 2021