Since 2010, we have been at the forefront of fully automated quantitative trading. Initially, we developed various prototypes of algorithmic trading based on traditional trading techniques such as indicators, chart patterns, and pure mathematical models. However, we found that these strategies would gradually lose their effectiveness after a certain period of time. We had to constantly adjust parameters or even develop new strategies.

Later on, we also adopted algorithms for high-frequency trading. The principle of high-frequency trading is simple: profit from the price differences between different exchanges and liquidity by buying low and selling high. It is essentially a competition of networks and physical infrastructure, rather than purely a trading technique. Although it deviated from our initial direction, it did bring us considerable profits.

Nevertheless, we never gave up on using market analysis and forecasting as the essence of our trading approach. Therefore, we have never stopped researching and testing trading strategies.

2016 was a remarkable year for us.

In 2011, IBM's artificial intelligence system Watson defeated human contestants on the American TV game show 'Jeopardy!'. In 2012, Google's deep learning algorithm achieved breakthrough results in the ImageNet Large Scale Visual Recognition Challenge, a competition for image recognition. In 2014, Microsoft's speech recognition system reached human-level performance and surpassed human participants in an international speech recognition competition. All these events indicated the immense potential of machine learning and deep learning in the field of computing.

Therefore, we also incorporated these two AI approaches into the core algorithms of our trading technology research and development. Additionally, to increase the sample data for learning, we shifted our focus from the securities market to the foreign exchange market.

In the same year, Google's artificial intelligence system AlphaGo defeated world champion Lee Sedol in a game of Go. This event was considered a significant breakthrough in the history of artificial intelligence, as Go is a highly complex game that requires a high level of intuition and strategic thinking. It further strengthened our conviction that researching artificial intelligence was the wisest choice, and it became a milestone for us.

In 2017, we independently developed the prototype of the Firefly series artificial intelligence deep learning platform.

Firefly I utilized trading algorithms based on technical indicators and historical data as the basis for trading strategies. However, the fundamental difference from previous approaches was the real-time recording of quote data and its utilization for dynamic parameter adjustments. That year, Firefly I brought us many surprises, with the highest annual return rate reaching an impressive 76.2%.

In that year, the research on manually selecting and identifying trading patterns also incorporated the use of artificial intelligence and deep learning for filtering. Our work transformed into merely defining which trading patterns are effective under specific market conditions. This saved us a significant amount of time and effort, allowing us to continuously research trading models.

Firefly II is a trading algorithm based on the logic of resistance and support levels in the long and short positions. It requires understanding the balanced distribution of long and short positions in the current trading instrument, and then analyzing the most probable direction and required strength of market fluctuations based on dynamic position analysis. Firefly II has been one of our crucial algorithms in maintaining a high success rate in recent years. It represents our core competitiveness.

In order to develop Firefly into an adaptive and continuously upgradable artificial intelligence trading platform, we devoted all our efforts to planning and developing a comprehensive trading ecosystem for Firefly. Moreover, that year, we were honored to receive the Fintech Nexus Awards 2019 for our achievement with Firefly II.

In order to achieve a complete independent system, we decided to move away from Refinitiv's aggregator and started using our in-house developed Firefly Spark Aggregator. This decision was made to better integrate our existing HFT Detector for enhanced efficiency. The Firefly system is structured into four layers: the Quote Layer, Trading Layer, Management and Scheduling Layer, and Clearing Layer. Over the course of two years, we have nearly completed the construction of the entire system, and we anticipate an official launch in the latter part of 2023.

During this period, we have been running internally for one year, achieving a remarkable performance with a peak return of 176.8%. This outcome has brought us a great sense of accomplishment.