Cindicator Bot: Progress to Date
In the past six months we have gathered more data than in the first three years of Cindicator’s existence. Since September, the number of data points we generate every month has increased by 30 times. We’ve organically increased the number of analysts on the collective forecasting platform by 10 times in less than half a year. During the last month, this number approached 100,000 as we started our marketing campaign. As the pace of our development accelerated, we deployed the first neural network to process more effectively the enormous volumes of data we are collecting.
Here are some additional statistics to illustrate the current scale:
- 400,000+ answers were gathered in March;
- 200+ indicators are sent every month;
- 20 machine-learning models are now active.
We invite you to learn more about our research and deepen your understanding of how indicators evolve.
Summary of indicators
Since the beginning of last year, we sent:
- 1,452 indicators in total;
- 1,298 binary indicators showing the probability of a financial event happening;
- 144 support and resistance level indicators;
- 10 ICO rankings (a new type of indicator added this year).
Here is a full list of all the indicators we’ve sent since Q4 2016. The spreadsheet starts with a recap of all indicators and their accuracy. It then outlines the breakdown of indicators sent in Q1 2018. It then lists the full information for all indicators sent during each previous quarter.
Now let’s review which indicators we sent in the first quarter of this year.
Relative to the previous period, token holders on the Expert tier have received the greatest number of indicators. Users on other tiers have also received a significant number of indicators. Users on the Explorer tier, which is above Beginner, have received fewer indicators relative to the other tiers because this tier was introduced in the middle of the quarter.
Price level indicators were the most common type of indicator in the first quarter. With each increasing tier, users receive exponentially more indicators in comparison to the number of tokens they must hold.
The number of indicators sent for crypto markets was almost four times greater than the number of indicators sent for traditional financial markets. We might add more fiat indicators in future, if that proves to be in line with market realities. We already have a number of indicators that worked well for EPS and could easily be turned into investment strategies.
What’s next: neural networks
As we recently announced, we now have our first fully functioning neural network. We’ve been experimenting with neural networks for a long time. We now finally have the volume of data required to reasonably capture complex relationships existing between different variables.
According to backtests, the new indicators based on the neural network show an average increase in accuracy of 21 percentage points even during periods of significant market turbulence. The backtest results of binary questions in Q1 are available in the same spreadsheet.
The backtests were done to retrospectively compare the accuracy of indicators with and without the neural network. All backtests were done with the same events that were used for real indicators. The neural network, of course, did not have any information about ‘future’ events.
All indicators sent after 22 March were already being processed by the neural network — 69% were accurate.
What drove this increase in accuracy?
The neural network adds a third learning layer in the process of creating indicators. First, analysts answer questions and learn with help of a motivational system. Our machine-learning models then assign different weights and corrections to their suggestions, depending on their past performance, combining collected data with other data from the web (currently it’s Twitter sentiment analysis — more web data sources are coming in Q2–Q4). Finally, the neural network locates complex, nonlinear dependencies between statistical models and gives an individual weighting to each model in order to arrive at the final indicator.
The Cindicator ecosystem creates a positive feedback loop. More analysts on the collective forecasting platform generate more data. This higher volume of data gives rise to smarter machine-learning models. In turn, these models, combined with the neural network, lead to more valuable products. Value creation directly (through larger reward funds) and indirectly fuels the growth of the collective forecasting platform.
The neural network learns from experience. The more data points the neural network has, the more accurate the indicators it produces. It’s great that the market conditions in the first quarter offered the Cindicator ecosystem plenty of unique opportunities to train Hybrid Intelligence and evolve.
Learning from volatility
Indicators learn to perform in any market and Hybrid Intelligence adapts to the new market climate. In the face of high market volatility and a persistent bearish trend, the neural network showed 66% accuracy, according to backtests.
Following a short period of sharp sell-off just before the New Year, the majority of cryptocurrencies were growing in the beginning of January again, contributing to the rapid increase in combined market capitalisation.
On 7 January, the combined market capitalisation of all cryptocurrencies climbed to over $830 billion. Less than a week later, on 13 January, Ether went above $1,400.
In January we also had a sharp downturn, which many crypto market participants misinterpreted as a healthy temporary correction that would shake out the weak hands. It turned out, however, to be a sustained downward trend. From the all time highs, the combined crypto market capitalisation declined 66.9% to $274 billion by 18 March.
While Bitcoin has previously experienced similar price swings, most other assets are too new and don’t have a comparable history. The entire first quarter of this year has been a bitter yet invaluable learning experience for all market participants.
Looking at a historical perspective of the market capitalisation and it’s corrections, we see that there were many 30+% corrections. The table below lists these corrections to date:
Adding just one neural network as a final layer in the entire ML pipeline had a major positive impact on the accuracy of indicators. It’s possible that other types of neural networks could yield a greater improvement. We envision having several neural networks that are constantly learning from ever increasing data sets.
We plan to add new models for other types of questions. Until recently, our models were mostly focused on increasing the accuracy of binary indicators. Adding more data scientists and establishing strategic partnerships with companies and institutes around the world will dramatically increase the pace of our progress.
For binary indicators, we plan to further strengthen the neural network by adding even more models. This would likely increase the accuracy of the neural network’s indicators. One interesting direction that we are considering is using a recurrent neural network as an ensemble model for price level indicators. As we continue to accumulate more data, the depth of learning and quality of indicators is set to increase due to the very nature of the neural network. We are also experimenting with the architecture of the whole system and its settings.
For support/resistance level indicators, we are aiming to develop a better understanding of the structure that would work best for model combinations. We expect several types of binary model to be adjusted for support/resistance levels and implemented into the production version of the pipeline. Moreover, we see great potential in the implementation of a neural network for this type of indicator as well.
Another area of our research is focused on models based on using time series in combination with technical analysis and high frequency real-time data. We also see great potential in real-time sentiment monitoring systems. Our data science department is currently exploring both types of models.
This has been a tough and very productive period. The neural network is learning rapidly. We are closely analysing every move of the market and implementing new models to increase the accuracy of our indicators. We believe that a clear and detailed view of the results of our work will give you more opportunities to make better decisions amid persistent volatility.
Thank you for your support!
The Cindicator team