Our mission at Neural Network Industries is to empower users with the capabilities of neural network trading algorithms. Our team of engineers and financial professionals envision to set up a disruptive business model with a focus on competitive durable performance as well as an developing an economic moat to maintain and build market share by delivering state-of-the-art trading algorithms.
At Neural Network Industries we are planning to provide ancillary services for our clients, for example constructing "botfolios" (Pareto optimized portfolios consisting of multiple bots operating in multiple trading pairs), API trading service as well as financial management. We are also currently in the formation of a legal entity, solely dedicated for the sale of our SLI token (an ERC-20 token called "SleipnirWallet") and the API trading commercialization. In our latest efforts to find suitable partners for the marketing of our API services and the sale of our SLI token, we have established several partnerships with marketing agencies in Singapore, Austria and Eastern Europe. The tentative commercialization of our API services will commence in 2020 and among our supported exchanges are Binance, Poloniex and Kraken.
In our current development stage we have established a groundbreaking trading strategy, which is utilizing an iterative optimization algorithm combined with a simple moving average indicator. Our two flagship bots deliver definitive results which can be monitored at https://acax.createaforum.com/governance/bot-performance-2019/ . Our scientific work in the field of artificial neural networks, finance and engineering has enabled us to create an array of neural network strategies, whereas one strategy yielded 108% return in the fiscal year 2018 (NNv2 USDT/BNB). Our two fully optimized trading algorithms bear the name "NNv2" and "NNv3" and are engineered to generate a sustainable return on investment during periods of high volatility and market unpredictability. In 2019 January we deployed the first neural network botfolio focused on the most liquid cryptocurrency pairs, consisting of five automated trading positions. Ultimately, we applied Pareto optimization to determine the best asset allocation in order to maximize the Sharpe ratio, which in other words is the excess return per unit of standard deviation.
Pareto optimization of 5 asset Botfolio
Botfolio Return in FY2019
The NNv2 trading algorithm has been trained with historical data of the digital asset Binance coin (BNB) since inception of BNB. It is usable in the USDT/BNB, USDC/BNB and PAX/BNB trading pair on Binance. NNv2 uses gradient descent to predicts price changes. NNv2 is based on the ADADELTA optimization algorithm, which was developed Matthew D. Zeiler (https://arxiv.org/pdf/1212.5701.pdf).
"Adadelta is an extension of Adagrad that seeks to reduce its aggressive, monotonically decreasing learning rate. Instead of accumulating all past squared gradients, Adadelta restricts the window of accumulated past gradients to some fixed size w." (Ruder.io , 2019)
Moreover, it features a "SMA SLOW" and "SMA FAST" indicator in order to detect bull and bear market periods by identifying SMA crossovers, for example when the "SMA SLOW" crosses above the "SMA FAST" it implies that the price momentum is increasing and vice versa. The optimal parameters of the algorithm (including the two SMA lengths) were optimized with genetic algorithms over 3 years training data of Binance coin. The NNv2 method has approximately 3% standard deviation per trade and 1% mean return per trade, while having 200 trades in a fiscal year. The code of the strategy is based off https://github.com/SirTificate/gekko-neuralnet.
SMA fast and SMA slow indicator
NNv2 PAX/BNB bot performance YTD
Let us now consider our next outstanding trading algorithm, which was build to be compatible with trading pairs across all asset classes. The NNv3 algorithm has been trained using 3–5 years training data of multiple digital assets (BTC, ETH, LTC, BCH, XRP, ETC, BNB) resulting in approximately 40 epochs trained per digital asset. We engineered two versions, the NNv3 STALWART version compatible with every trading pair and the NNv3 BNB version, solely optimized to deliver uncorrelated (negative correlation with the BNB market and the NNV2 trades) botfolio performance, with the objective to outperform the BNB market. The NNv3 STALWART neural network strategy is compatible with most cryptocurrency trading pairs, due to its defensive risk-preference. Moreover, were this strategy's backtesting results characterized by a good winning ratio of 2.22 (68.9% winning percentage), low amount of trades (approx. 30 trades in a year), 4%-8% standard deviation and 1%-5% mean return across all aforementioned trading pairs. This algorithm is optimized with the objective to achieve risk-adjusted return during market periods of uncertainty or high volatility. NNv3 has a proven profitable track record and is engineered to execute high probability scalp trades during extreme market drawdowns. We have automated two NNv3 trade algorithms to be broadcasting their trades on Twitter (https://twitter.com/neuralnetworkin) and Stocktwits (https://stocktwits.com/NeuralNetworkIndustries) using USDC/ETH and USDC/BTC trading pairs quoted on Binance.
NNv3 Stalwart USDC/BTC bot performance YTD 07/2019
NNv3 Stalwart USDC/ETH bot performance YTD 07/2019
NNv3 USDC/BNB bot performance YTD 07/2019
Continuous improvement and quality assurance for our strategies is a top priority for our team. We have established a repertoire of cross-validation and regularization techniques in order to cope for over-fitting. We also have determined a suggested position size for the various trading algorithms in accordance with daily volume, track record and slippage (800-1000 USD principal in one trading position is the upper threshold recommended). Additionally, sample data designated for training of the algorithm has to be distinguished between two types: in-sample and out-of-sample data. The in-sample performance has been simulated over the sample used in the design process of the strategy (training data). The out-of-sample performance is simulated over the live data (test data). A general rule of thumb in back-testing theory is that a successful strategy should yield similar performance in test data as well as training data, which applies to our NNv3 strategy. Other attributes of NNv3 is that it yields similar results across multiple digital assets. Our conceptual framework for continuous optimization of the algorithm will be facilitated by weekly walk-forward testing on multiple digital assets out-of-sample data.
Stay up-to-date with News at: https://neuralnetworkindustries.blogspot.com/
Check out our whitepaper: https://acax.createaforum.com/botconomy/whitepaper-rough-draft-v0-8/
Find out more at: https://acax.createaforum.com/
Follow us in Twitter: https://twitter.com/neuralnetworkin
Follow us in Stocktwits: https://stocktwits.com/NeuralNetworkIndustries/
No comments:
Post a Comment