Combining different sentiments can pay off: Based on Augmento’s sentiments we have built sentiment polarity indicators that exhibit correlations of up to 0.5 (Spearman) with forward returns. You will learn how to combine different sentiments into a single indicator and how such indicators are dependent on underlying sentiments, current market cycles, and time horizons.

At Augmento we quantify more than 100 sentiments and topics in discussions about cryptocurrencies on social media—Twitter, Bitcontalk and Reddit—in real time.

Some of these sentiments are polar opposites:

  • Positive vs. Negative
  • Good_news vs. Bad_news
  • Bullish vs. Bearish

The SentPol Indicators are built upon SMAs but based on selected sentiments from Augmento’s AI.

Strength and direction of the relationship between the indicators and forward returns of the selected cryptocurrencies depend on multiple parameters, including forward horizon (f) of the return and size of the MA window (w) used in the simple moving averages. The SentPol Indicators are calculated as follows:

Where S+/– are opposing sentiments, ΔS+/- denotes the change of S+/- since the previous day, and SMA is the simple moving average with window w. The following formula is used to calculate forward returns of price P at time t, for f days forward:

The correlations between SentPol and forward returns of certain crypto assets reach up to 0.5 (Spearman). Our findings indicate that synthesizing indicators based on different combinations of Augmento sentiments/topics can be leveraged in different trading contexts.

Short term vs long term

Some indicators first experience a sharp rise of correlation with forward returns followed by long tails of declining correlation.

Fig. 1: Correlations between SentPol based on positive/negative sentiments towards Ripple and forward returns of Ripple in 2017

Other indicators initially lay flat and then gain significant relevance on a longer time horizon.

Fig. 2: Correlations between SentPol based on Good news/Bad news towards Ripple and forward returns of Ripple in 2017

Sometimes: counterintuitive logic

Some sentiment indicators follow an—at first sight—counterintuitive logic. For example, indicators based on bullish and bearish sentiment often start in an area around zero to then develop a strong negative correlation with forward returns (Fig.3 and Fig.4)—a finding which could be harnessed for counter trading strategies.

Fig. 3: Correlations between SentPol based on Good news/Bad news towards Litecoin and forward returns of Litecoin in 2018

Fig. 4: Correlations between SentPol based on Good news/Bad news discussions towards Bitcoin and forward returns of Bitcoin in 2018

Market cycles matter

Finally, the findings indicate that sentiments/topics need to be interpreted in relation to the associated market cycle. For example, an indicator based on discussions around Good_news and Bad_news for Ethereum were positively correlated in a bull market, while exposing an opposite trend in the bear market of 2018 (Fig. 5 and Fig. 6).

Fig. 5: Correlations between SentPol based on Good news/Bad news discussions towards Ethereum and forward returns of Ethereum in 2017.

Fig. 6: Correlations between SentPol based on Good news/Bad news discussions towards Ethereum and forward returns of Ethereum in 2018.

Statistical relevance

P-values are dependent on selected MA-timeframes, underlying sentiments and horizon of forward returns. We deem the results as statistically significant with most p-values being extremely small (such as: coin = ripple, datasource = twitter, sentiments = Positive/Negative, forward return horizon = 12 days, MA window = 7 days with a p-value of 1.048335 e-12) while some p-values are bigger but still close to  zero (such as: coin = ethereum, datasource = twitter, sentiments = Positive/Negative, forward return horizon = 14 days, MA window = 16 days with a p-value of 0.000188).

Sample size of the evaluated data is 730 (number of days of 2017-2018).

Caveats and further work

The next steps for research on the SentPol indicators are to perform this analysis with hourly sentiment and pricing data instead of daily data. The correlations above show that there is a strong signal in this data and that using combinations of sentiments and topics may pay off. However, the correlations often gain relevance after a couple of days which can have implications in live trading contexts, for example in times of high volatility. Further, the statistical significance of our findings are reduced to some extent due to the low frequency and relatively small sample size of the data. Any additional work should also include analysis with higher frequency data.

Leave a Reply