HYU-PaulKim / ML_Summary

brief summary of what I've learned so far from univ.
0 stars 0 forks source link

4. Conclusion #5

Open HYU-PaulKim opened 3 months ago

HYU-PaulKim commented 3 months ago

Conclusion for each approach

HYU-PaulKim commented 3 months ago

1. Correlation between BTC and other cryptos:

No cryptocurrency showed a time lagged cross correlation higher than the Pearson correlation coefficient or the Spearman correlation coefficient. This indicates that it is reasonable to consider that Bitcoin's fluctuations immediately affect other cryptocurrencies. Cryptocurrency prices showed a significant correlation with the price of Bitcoin regardless of whether outliers were removed. This suggests that diversified investment is not meaningful. Notably, meme coins such as Dogecoin and Shiba Inu Coin had higher outlier rates and differences in Pearson correlation coefficients than other coins. This indicates that memes have a significant impact on prices rather than the utility of the coin itself.

2. Exercise ML models to explain BTC price by on-chain data

2-A:

Since hashrate represents the computational power needed to mine Bitcoin, it is evident that there is a high correlation with difficulty. It seems natural that both individuals holding between 0.01 and 1 Bitcoin and those holding between 1 and 10 Bitcoins are considered "small investors," which explains the high correlation observed. Interestingly, there is a very strong correlation between hashrate and the number of "small investor wallets." This could be interpreted as miners getting more involved in mining as the "market attention," represented by small investors, increases. Another intriguing point is the clear negative correlation between the number of wallets holding 100 to 1,000 Bitcoins and those holding more than 1,000 Bitcoins. There are two main explanations for this: First, those holding more than 1,000 Bitcoins have a significant influence on the market. If they cannot dominate the market themselves, they are more likely to be influenced by the market compared to small investors. Alternatively, it is possible that individuals holding 1,000 Bitcoins are distributing them across 10 different wallets for risk management or inheritance purposes. In such cases, the strong negative correlation becomes quite clear.

2-B:

Subsequently, I attempted machine learning based on the remaining on-chain data, but none of the models outperformed the control group and, in most cases, performed worse than the control group. RandomForest, which showed performance close to the control group, essentially gave equal weight to all variables, while ElasticNet excluded the influence of all variables. This suggests that the provided on-chain data itself is not significantly related to Bitcoin price fluctuations.

3:

Bitcoin price fluctuations showed a positive correlation with Nasdaq, SOX, and SP500. Given that there was virtually no correlation with gold (XAU), it indicates that Bitcoin is treated more as a "risky asset" rather than gold. Although there is a strong negative correlation with KOSPI, it is very interesting to note that a strong positive correlation appears when there is a two-day lag. While it is difficult to pinpoint a clear reason for this, consider that Bitcoin and the Korean stock market are more sensitive to certain issues compared to other assets. It is possible that KOSPI might follow Bitcoin, which trades 24/7.