Forecasting - Drift MethodIntroduction
Nothing fancy in terms of code, take this post as an educational post where i provide information rather than an useful tool.
Time-Series Forecasting And The Drift Method
In time-series analysis one can use many many forecasting methods, some share similarities but they can all by classified in groups and sub-groups, the drift method is a forecasting method that unlike averages/naive methods does not have a constant (flat) forecast, instead the drift method can increase or decrease over time, this is why its a great method when it comes to forecasting linear trends.
Basically a drift forecast is like a linear extrapolation, first you take the first and last point of your data and draw a line between those points, extend this line into the future and you have a forecast, thats pretty much it.
One of the advantage of this method is first its simplicity, everyone could do it by hand without any mathematical calculations, then its ability to be non-conservative, conservative methods involve methods that fit the data very well such as linear/non-linear regression that best fit a curve to the data using the method of least-squares, those methods take into consideration all the data points, however the drift method only care about the first and last point.
Understanding Bias And Variance
In order to follow with the ability of methods to be non-conservative i want to introduce the concept of bias and variance, which are essentials in time-series analysis and machine learning.
First lets talk about training a model, when forecasting a time-series we can divide our data set in two, the first part being the training set and the second one the testing set. In the training set we fit a model to the training data, for example :
We use 200 data points, we split this set in two sets, the first one is for training which is in blue, and the other one for testing which is in green.
Basically the Bias is related to how well a forecasting model fit the training set, while the variance is related to how well the model fit the testing set. In our case we can see that the drift line does not fit the training set very well, it is then said to have high bias. If we check the testing set :
We can see that it does not fit the testing set very well, so the model is said to have high variance. It can be better to talk of bias and variance when using regression, but i think you get it. This is an important concept in machine learning, you'll often see the term "overfitting" which relate to a model fitting the training set really well, those models have a low to no bias, however when it comes to testing they don't fit well at all, they have high variance.
Conclusion On The Drift Method
The drift method is good at forecasting linear trends, and thats all...you see, when forecasting financial data you need models that are able to capture the complexity of the price structure as well as being robust to noise and outliers, the drift method isn't able to capture such complexity, its not a super smart method, same goes for linear regression. This is why more peoples are switching to more advanced models such a neural networks that can sometimes capture such complexity and return decent results.
So this method might not be the best but if you like lines then here you go.
Forecastingtechniques
Alpha-Sutte ModelThe Alpha-Sutte model is an ongoing project run by Ansari Saleh Ahmar, a lecturer and researcher at Universitas Negeri Makassar in Indonesia, that attempts to make forecasts for time series like how Arima and Holt-Winters models do. Currently Ahmar and his team have conducted research and published papers comparing the efficacy of the Alpha-Sutte and other models, such as Arima and Holt-Winters, on topics ranging from forecasting Turkey's CPI data, Bitcoin prices, Apple's stock prices, primary energy supply of Indonesia, to infant mortality rates in China.
The Alpha-Sutte model in comparison to the other two models listed above shows promise in providing a more accurate forecast, and the project has been able to receive some of its funding from organizations such as the US Agency for International Development, which is a part of the US Federal Government, so maybe the project has some actual merit.
How it works:
In this model there are four values presented at the top of the window.
1) The first value in blue is the value of the Alpha-Sutte model whose purpose is to forecast the price of the current bar.
2) The second value in yellow is an adaptive version of the Alpha-Sutte model that I made. The purpose of the adaptive Alpha-Sutte model is to expand upon the Alpha-Sutte by allowing new information to be introduced, causing the value to change during the current period, hence the adaptiveness of it.
3) The third value in aqua is the moving average of the low% Sutte line which is a predictive line that is based off of the close and low of the current and previous periods.
4) The fourth value in red is the moving average of the high% Sutte line which is a predictive line that is based off of the close and high of the current and previous periods.
Trend signals:
If low% Sutte (aqua value/line) is greater than high% Sutte (red value/line) then this is a buy signal.
If high% Sutte (red value/line) is greater than low% Sutte (aqua value/line) then this is a sell signal.
Caveat:
Even though this model's purpose is to forecast the future, will it be able to predict periods of large movements? No, of course not, but it will adjust quickly to try to make more accurate forecasts for the next period. This was also a reason why I made an adaptive version of this model to try to reduce some of the discrepancies between the Alpha Sutte and price when there is a large unexpected move.
*WARNING before using this I would highly recommend that you look up "Sutte Indicator" online and read some of the papers about this model before you use this , even though this model has shown merit when compared to Arima and Holt-Winter models this is still an ongoing project.*
Hopefully this project will actually come to something in the near future as the calculation for this time series predictive model is much easier to calculate and program in pine editor than something like an Arima model.
*Also, if you know how to use R language there is a package for the "Alpha-Sutte model".*
Linear ExtrapolationBasic extrapolator for forecast a time-series, all forecasts are mades length periods ahead.
This is not a estimation of the exact price
This should only be used for forecasting direction, dont expect the price to be at the same value of its forecast.
Bias, Mean absolute error, Mean percentage error...etc look useless here, its better to use correlation as a accuracy measurement.
Correlation(Forecast ,close,period)
Rescaling for a better forecast ?
Transforming a non-stationary signal to a stationary signal can increase the forecasting accuracy, this can be done by detrending. Here is a list of somes detrending methods:
Auto-Bias : price - price
Mean-Bias : price - price moving average
Log transform : log(price/price moving average)
Correlation : correlation(price,n,period)