Results
Last updated
Last updated
The models were trained with the optimized parameters provided in the respective papers. The accuracy of our model has been found to be outstanding. You can find the training and testing metrics for the referenced papers in Table 5. The classification accuracy of our model, considered as a model predicting continuous variables, outperforms our baseline benchmark. The model exhibits better recall and accuracy, as clearly demonstrated by the f1-score compared to the two reports by Kara Y et al. (2011) and Hoseinzade E et al. (2019). The LSTM model considers data from the past 30 days and captures patterns before making predictions. While the CNN convolutions view the numbers as static images, the output results will be the same if the numbers are identical, even in increasing or decreasing trends. The LSTM model built upon the CNN model helps address this issue and thus makes the model more intelligent:
Metrics
Train - Test
Train - Test
Train - Test
F1-score
0.046 - 0.16
0.009 - 0.07
0.575 - 0.581
Accuracy
0.473 - 0.496
0.382 - 0.366
0.539 - 0.533
The performance of our model on the training and testing sets can be observed below in Figures 6 and 7, respectively.
The model in this report is compared to traditional buy-and-hold strategies in the Spot market and the Futures market of Solana over a period of 1 year. The traditional buy-and-hold method incurred a loss of -88% from January 1, 2022, to February 14, 2023. In contrast, our model in the Futures market with automated TP/SL optimization generated a profit of 288% during the same testing period. This model yielded significantly higher profits compared to the traditional buy-and-hold strategy.
Backtesting period
Jan 2022 - Jan 2023
Beginning Balance
$5,000
No. trades
63
% largest profitable trade
134.99%
% largest loss trade
-3.27%
No. win trades
38
Win rate
60.32%
Loss rate
39.68%
Average profit / capital
7.85%
Average loss / capital
-0.38%
Profit Factor
20.77
Ending Balance
$19,440
Ending Profit
$14,440
Profit Margin
288.80%
Average holding day(s) of loss trade
2.5
Average holding day(s) of win trade
8.9
The data provided outlines the effectiveness of trading based on signals for SOLUSDTPERP, a futures asset traded on Binance, from January 2022 to January 2023. The initial investment capital was $5,000, and a total of 63 trades were executed. The highest profit rate of winning trades was 134.99%, while the highest loss rate of losing trades was -3.27%. The win rate was 60.32%, and the loss rate was 39.68%. On average, the trades generated an average profit of 7.85% on the initial capital, with an average loss of -0.38%. The Profit factor was 20.77, indicating significantly higher total profits compared to total losses. The ending capital was $19,440, resulting in a final profit of $14,440. The impressive return on initial capital reached 288.80%. The average holding time for losing trades was 2.5 days, while the average holding time for winning trades was nearly 9 days. The relatively short average holding time for losing trades indicates that the trading system has effectively identified stop-loss positions to limit losses. Meanwhile, the longer average holding time for winning trades demonstrates that the system has been able to identify and capitalize on favorable market conditions to maximize profits. Overall, the data shows that the signal-based trading strategy in the Futures market is effective in generating profits by leveraging market opportunities while effectively managing take-profit/stop-loss (TP/SL) through effective risk management.
