-
Pure Gamma Trading
PG Research was founded in 2017 and has been involved in the Crypto market, quietly trading, monitoring public opinion and generating revenue in various communities since its inception.
The information presented on this web site is not a solicitation for investment. Such investment is only offered on the basis of information and representations made in the appropriate offering documentation. Past performance is not necessarily indicative of future results. No representation is being made that any investor will or is likely to achieve similar results. Crypto trading is speculative, involves substantial risk and is not suitable for all investors.
Read about the team composition, background, and past performance, along with financial algorithm research, as well as our framework: https://docsend.com/v/94csc/cn
We are a $500 million portfolio management team, including $20 million in proprietary capital. Our team members often work on market structure and data engineering from unconventional locations like the beach or home offices. If that suits your style, we’d welcome the opportunity to collaborate.
https://www.rootdata.com/Investors/detail/PG%20Trading?k=MTI3NDk%3D
Nautilus – from ancient Greek ‘sailor’ and naus ‘ship’.
The nautilus shell consists of modular chambers with a growth factor which approximates a logarithmic spiral. The idea is that this can be translated to the aesthetics of design and architecture.
Abstract To address the non-stationarity and high-noise characteristics of order flow in highly liquid cryptocurrency assets (BTC/USDT), this study proposes a multi-scale signal generation framework integrating Continuous Wavelet Transform (CWT), LSTM-ARIMA hybrid model, and Relative Strength Index (RSI). By decomposing 2-7 days of historical order flow data (price, volume, order book depth) using CWT, mid-to-short-term components (1h-6h) are extracted to suppress high-frequency noise, while low-frequency components (>6h) are input into an ARIMA model to analyze mean reversion and seasonal fluctuations. High-frequency components (<1h) are processed by an LSTM network to capture nonlinear dynamics such as market sentiment shifts and liquidity pulses. The RSI indicator is introduced to optimize signal trigger thresholds, and cointegration error filtering (BTC-BNB spread portfolio) is applied to eliminate anomalous signals. Empirical analysis based on 2023 BTC/USDT high-frequency data from Binance demonstrates that the generated signals achieve a Sharpe ratio of 3.72 (a 26.4% improvement over the standalone LSTM model), maximum drawdown of 1.9%, and a win rate of 58.7%, validating the model’s effectiveness in noise reduction, dynamic response, and risk control.
Model Architecture and Empirical Analysis Multi-Scale Decomposition and Feature Extraction Utilizing the Morlet wavelet basis function, the order flow data undergoes CWT decomposition to isolate high-frequency noise (<1h), mid-term trends (1h-6h), and long-term cycles (>6h). Low-frequency components are modeled via ARIMA to capture linear statistical patterns, while high-frequency components are fed into an LSTM network to detect nonlinear dynamics such as liquidity pulses and “whale address” anomalies.
Signal Fusion and Optimization A dynamic weighting mechanism integrates signals from LSTM (nonlinear responses) and ARIMA (steady-state trends), augmented by RSI (14-period) to filter overbought/oversold signals and reduce false trading frequency. Cointegration filtering constructs a BTC-BNB spread portfolio, leveraging the Johansen test to establish long-term equilibrium relationships and dynamically discard signals deviating from the cointegration space.
https://github.com/HaoranXue/Machine_Learning_For_Structured_Data
<iframe src="https://docsend.com/v/94csc/cn" allow="fullscreen" width="640" height="480"></iframe>