R&D records in MIRLab

# R&D records in MIRLab ## Reference papers ### 1. Convolution on neural networks for high-frequency trend prediction of cryptocurrency exchange rates using technical indicators * [Paper Link](https://www.sciencedirect.com/science/article/abs/pii/S0957417420300750) * [Dataset](https://) * [Code in GitHub(Not organized yet)](https://) * The accuracy of forecasting the trend of tomorrow's bitcoin price is over 60% in this paper by using c-lstm model. * Use 10 kinds of technical indicator as the features. * I try to implement the algorithm and the accuracy of forecasting the trend of tomorrow’s bitcoin price is about 53% in my test. Need to figure out some problems on the following to improve accuracy: * The size in covn2d kernel of c-lstm model is 1x1, but the 1x1 convolution kernel is only used to adjust the dimension. Would that omit to calculate the relation between the technical indicators and the price per minute? * I think the loss function should be crossentropy, but the output layer is 1x1 not 2x1. * Had the author normalized the data before entering the model? If they had, did they use the StandardScaler? * Did the author flatten the feature before sending it into LSTM or keep the shape of 15(time-step)*18(indicator)? ### 2. Bitcoin Price Forecasting via Ensemble-based LSTM Deep Learning Networks * [Paper Link](https://ieeexplore.ieee.org/abstract/document/9333853/metrics#metrics) * [Dataset](https://) * [Code in GitHub(Not organized yet)](https://) * Use three different time interval LSTM NN model(3-min, hour and day) and then composite to an ensemble model. * The accuracy of the ensemble model is better than three LSTM NN model on Bitcoin forecasting case in this paper. * I compare different conditions about using the same ensemble model, and the result is on the following: * is being processed * There are some issues that I think need to be carefully considered: * The calculation method of accuracy will make the result close to 100%, and the predicted result looks more like shifting the actual price backward. * Although the result of the Ensemble LSTM NN model is better than the other three LSTM NN models (minutes/hours/days), the Ensemble model does not cover more information, only because of the linear optimization process. ### 3. Deep learning in predicting cryptocurrency volatility * [Paper Link](https://www.sciencedirect.com/science/article/abs/pii/S0378437122001704) * Consider two types of Neural Network suitable for time series: the NLANN and the JNN(Jordan Neural Network). * Although the predicted volatility results are better than other models, there is still a delay. ### 4. Cryptocurrency Portfolio Management with Deep Reinforcement Learning * [Paper Link](https://paperswithcode.com/paper/cryptocurrency-portfolio-management-with-deep) * Comparing the difference in the performance of CNN models and other recently published portfolio selection strategies. * The reward function use the total capital change after each trading period to calculate, and do not use sharpe ratio and maximum drawdown. * The return of the portfolio using CNN model is worse than 'Passive Aggressive Mean Reversion' method. ### 5. Pump and Dumps in the Bitcoin Era: Real Time Detection of Cryptocurrency Market Manipulations * [Paper Link](https://paperswithcode.com/paper/pump-and-dumps-in-the-bitcoin-era-real-time) * Study the relationship that exists between the groups, the exchange, and the target cryptocurrencies. * Present a tool that can detect in real time a pump and dump in action. ### 6. KryptoOracle: A Real-Time Cryptocurrency Price Prediction Platform Using Twitter Sentiments * [Paper Link](https://paperswithcode.com/paper/kryptooracle-a-real-time-cryptocurrency-price) * The tweets extracted from Twitter are cleaned of non-alphanumeric symbols and then processed with VADER (Valence Aware Dictionary and sEntiment Reasoner). * The compound sentiment score is then used to establish correlation with the Bitcoin prices over different lag intervals. * It builds a prediction engine that forecasts Bitcoin prices at specified intervals. ### 7. Stock Selection via Spatiotemporal Hypergraph Attention Network: A Learning to Rank Approach * [Paper Link](https://ojs.aaai.org/index.php/AAAI/article/view/16127) * ### 8. High-frequency Statistical Arbitrage Strategy Based on Stationarized Order Flow Imbalance * [Paper Link](https://www.sciencedirect.com/science/article/pii/S1877050921008917) * ### 9. DSLOB: A Synthetic Limit Order Book Dataset for Benchmarking Forecasting Algorithms under Distributional Shift * [Paper Link](https://arxiv.org/pdf/2211.11513v1.pdf) * ### 10. Encoding of high-frequency order information and prediction of short-term stock price by deep learning * [Paper Link](https://www.tandfonline.com/doi/abs/10.1080/14697688.2019.1622314) * ### 11. The life of U’s: Order revisions on NASDAQ * [Paper Link](https://www.sciencedirect.com/science/article/pii/S0378426619302973) *