What would the future of investing be like in the next decade?
Many of us doesn’t realize how much algorithms plays a role in our daily lives. From the contents that shown in your screens to the millions of bots that runs the show in wall-street. Should we hand over the baton to the machines to do our market analysis? Is it even humanly possible to make a fast and precise decision when trading in the cryptocurrency market? This is where KepingAI Long-Short Time-Series Forecasting comes in. We believe that pre-determined investing strategies paired with the power of deep learning models is the future of investing.
KepingAI LSTF is a cluster of multi-horizon forecasting models that predicts the price spread in percentage and direction of multiple future time steps of many cryptocurrency markets. Instead of the typical one-step ahead forecast mechanism, KepingAI LSTF forecasts an approximation of the price at the 8th future time step. Allowing the models to optimize trading actions at multiple time-steps in the future for 50+ cryptocurrency markets in parallel. Each model is fine-tuned to a specific cryptocurrency market with approximately 384 million machine learning parameters. Currently we are offering 50+ models trained and maintained to specific cryptocurrency market deployed in our layer 2 cluster, which we call the intelligence engine.
Layer 1–2
Signal & Intelligence Engine
Signal Engine runs in an orchestration architecture that retrieves real-time market data from multiple exchanges at once to scout for potential actionable intelligence from a specific cryptocurrency market. We use static-algorithmic strategy to scout for potential market volatility in parallel before sending the processed market data to layer 2. Layer 1 uses standard OHLCV data from exchanges web sockets with an average latency of 0.05s. This market data are then converted into momentum, volatility, overlap indicators, and more to run technical analysis in real-time.
Currently, we are using the 15m interval for several reasons but mainly due to the trading strategy that we optimize for focuses on high-turnovers. We monitors 50+ cryptocurrency markets consists of the major top 20 cryptocurrency based upon their market caps to ensure liquidity and prevention of slippage on our layer 3 (execution engine). Our models are trained on quote instrument USDT and will be expanding our horizon to other quote instruments such as BTC, ETH & BNB.
The signal mechanism in this layer prevents any duplications, i.e. there can only be one signal per ticker with a vertical barrier (time-expiration) of 8 hours (32 tick interval to the future). Layer 1 is independent to layer 2 for which it only generates pure signal based upon the given algorithmic technical analysis currently running. In addition, layer 1 monitors for both long & short market volatility.
KepingAI LSTF auto-regressive pre-trained model consists of x million machine learning parameters which are trained with static co-variates of different cryptocurrency nominals. In this pre-training model, we decided to use the highest market-cap of a cryptocurrency market for each nominal variates. The goal of this pre-train model is to have a generic well-balanced model that understands major cryptocurrency market behaviors fo the last 5 years. We uses fractional, ones, tens, hundreds, thousands, and ten-thousands variates to enable scalability for future fine-tuning use cases. This pre-trained model are then re-trained manually every quarter to ensure it’s validity and awareness of recent changes of the market. The main purpose of this pre-train model is to prevent the off-chances of overfitting on the next step of our training mechanism by training the heavy load in the beginning.
KepingAI LSTF utilizes state-of-the-art multi-horizon time series forecasting model architecture for predicting the spread and direction of a given cryptocurrency market. It has approximately 384million machine learning parameters, with an average of 0.8+ precision for forecasting an approximation of both long and short signals. This fine-tune model is trained to optimize high-turnovers for signal users by ensuring its ability to forecast and approximation of the price and spread with very small error rate. The average duration for each signal generated from this fine-tune model ranges from 30–40minutes to completion (purchase and exit with a profit).
Our discovery of KepingAI LSTF models gives us a clue as to what may be the general deep learning model adoption to time-series forecasting use cases. With the proper maintaining architecture and dedicated iterative research team, KepingAI LSTF manages to maintain a very high-precision performance with its ability to be scaled as needed.
Layer 3
Execution Engine
We’ve paired our LSTF models with KepingAI’s Dollar-Cost-Average Bots. Users in our platform are able to configure their bots in a personalized way that matches their risk appetites. To ensure that the DCABots behaves symbiotically with our models, configurations such as take-profit, stop-loss, timeouts and safety-orders are pre-determined algorithmically per signal generated. These configurations dynamically change depending on the market conditions and the aggressiveness of our models. This allows us to maximize risk management and profitability for each trade without having the need for manual intervention.
The sole mission of this layer is to ensure the self-sustainability of our models while bringing generous yields for our users. This is achieved by having dollar-cost-average as our risk management system to ensure that false-positives predictions can be mitigated.
Click here if you’re interested in KepingAI Trading Platform
Click here if you’re interested in learning more about our models.
