What is the point of my "Chinese Slumber Method" predictions?
I do believe that the price of bitcoin is basically unpredictable, apart from it following the "geometric Brownian" model with variance modulated by trade volume. That is, the market does not care about past prices, only about the current price; hence it has no "trends" based on the price itself.
However, I thought that there may be underlying trends that persist for several days, due to "external" variables such as the amount of money and bitcoins in the exchange, the exit of old investors and recruitment of new ones, and persistent trader sentiment modified by news and price evolution. I had the hunch that those trends would be best seen by considering the price at Huobi at some fixed time in the early morning.
Why Huobi? First, until recently it was the largest exchange by volume, and thus more likely to influence other exchanges than be influenced by them. Second, it is a simple market: one product (BTC), one currency (Yuan), input and output of currency only through Chinese banks, and a clientele largely confined to China and so presumably more uniform that that of other exchanges. Third, because its clieants all but stopped trading at small hours of the night.
Why look at Huobi's price in the early morning? Because, looking at the charts, it seemed to me that the price at those times followed simple trends, over several days, whereas the price during the day often deviated from such trends. Asa possible explanation, I though that perhaps many traders tried to return to their preferred BTC:CNY positions before going to bed, after having deviated from those positions during the day; and that somehow brought the price back to some "natural price" determined by those hypothetical exeternal variables.
The daily predictions are a crude attempt to verify this impresion. I choose 19:00--19:59 UTC (03:00-03:59 China time) as the sampling hour ("Slumber Time"), since volume was minimum around that hour. I picked a threshold volume at that hour (500 BTC at first, then 0.5% of the total daily volume) to exclude data points when trade continued well into the night (the "Abertosaur" points). Again by looking at the charts, it seemed to me that, on those "long days", the price often deviated from the trend line even at slumber time. Moreover, breaks in the trend line seeemed to occur mostly on those days. Then I tried to fit simple trend formulas to strings of consecutive "good" data points (the "Glyptodons"); either continuing across the bad ones, or breaking at them.
For trends spanning a few days, it does not make much difference to work in linear or log scale. The general trend formula I use is the shifted exponential P(d) = A + B*Q**(d-d0), where A, B and Q are parameters to be determined, d is the day of the month, and d0 an arbitrary starting day. That formula can be used when there are at least three data points; the value of Q can be estimated from the price differences between consecutive days, and then A and B can be fitted by the least squares method. A special limiting case of that formula is the straight line P(d) = A + B*(d-d0). The justification for that formula is that it is the expected behavior of many simple sistems, such as a reservoir where the input flow is constant and the output flow is proportional to the water level. (Think of new investors joining, and traders becoming discouraged and leaving, for example.)
This approach is not very "scientific" of course. A more serious analysis of the idea should, for starters, use a weighted average of the prices and volumes during the night, say from 18:00 to 20:59 UTC. Instead of classifying the data points into "good" and "bad" by an arbitrary threshold, one should map the slumber-time volume to a numeric "slumber weight", 1.00 for days of near-zero volume and tending gradually to 0.00 as the volume increases beyond that threshold. Then one should use weighted least squares to fit the trend lines, and dynamic programming to choose the most likely breaks. However, I don't know enough statistics to determine whether the trends obtained by such a complicated process are statistically significant or not.
Anyway, it does not seem to be worth going to all that trouble, considering that Huobi may goxify at any moment...