Greetings --
AmiBroker, and all the other charting and trading system development platforms, were designed for the specific purpose of reading, displaying, and manipulating financial data, and using a built-in model (decision tree) to interpret signals, generate buy and sell signals, prepare reports, and interface with brokerage applications.
Python is a general purpose programming language. In order to generate trading signals, the support that is built-in to AmiBroker must be programmed into Python, creating a development platform.
What we give up in moving from a traditional trading system development platform is all of the built-in applications. What we gain is the ability to expand beyond decision tree models. Whether that is worth while depends on how you use AmiBroker and how you want to use Python.
You will need to provide the features and capabilities you currently use in AmiBroker in Python. But you will not need to replicate all of AmiBroker if you do not use all of AmiBroker.
Do not misinterpret my comments. I think AmiBroker is the best of the traditional trading system development platforms. For traders who use the graphical capabilities and / or are satisfied with decision tree models, there is no need to leave AmiBroker for Python.
The reason to use Python is that it provides the capability to use models other than decision trees, and also to do other calculations in the same program that generates the trades.
Chapter 8, Model Development - Machine Learning, of my "Quantitative Technical Analysis" book, has fully disclosed and ready to run examples of the following sequence of programs and explanation:
1. Use AmiBroker to read primary data for a tradable issue, compute the variable-lookback version of the RSI indicator, generate impulse Buy and Sell signals from that indicator for trading Market-on-Close, compute and plot the cumulative equity. Begin on page 274.
2. Provide a flowchart, discussion, and template code necessary to implement a computer program that replicates the capabilities of the AmiBroker program in step 1. Begin on page 278.
3. Use Python to implement a program that replicates the AmiBroker program from step 1. This is a very bare-bones trading system development platform. Begin on page 279.
4. Use Python and state signals to do the same, verifying that state signals are equivalent to impulse signals. Begin on page 283.
The AmiBroker function for the program in step 1 to implement the variable-lookback RSI is about 20 lines. The block of control variables is another 12 lines. The code needed to set up the data, assign values to indicators, compute signals, generate trades is about 7 lines. Plotting takes another 5 lines. The whole program is a page and a half.
The Python program in step 4 requires more code in order to implement those functions built-in to AmiBroker. Loading the libraries is about 8 lines. Computing the custom RSI is about 25 lines. Loading the data and setting the date range is about 20 lines. Computing the indicator and interpreting to generate Buy and Sell signals is about 20 lines. The loop that interprets the Buy and Sell signals, generates the trades, and manages the account balances is about 50 lines. The whole program is about four and a half pages.
Moving on to machine learning.
Everything done above uses decision tree model. The Python program in step 4 illustrates a very basic decision tree. To use a different model, begin on page 365. Several functions that might be useful have been coded in Python and are included in the listing. Variable-lookback RSI, RateOfChange, z-score, softmax, DetrendedPriceOscillator, NumberZeros, GainAhead, AverageTrueRange, and PriceChange. Even when the documentation is as long as the code, each function is about 30 lines long. Data is read on-the-fly from Quandl. The process of bringing previous values forward so that each day is a complete and independent data point is illustrated. Data is divided into in-sample and out-of-sample, and target is isolated from predictor variables, in about 10 lines. The Logistic Regression model is fit to the in-sample data in one line. And the out-of-sample data tested in one line. Using more code than necessary, the confusion matrix showing the accuracy of the model uses about 30 lines (it could be one line).
Additional processing includes one line to plot, one line to compute terminal equity, one line to implement Monte Carlo analysis. Continuing on, if you wish, one page to implement dynamic position sizing.
Caveats --
What I have written and published is bare-bones, and is a template. You will be able to switch from Logistic Regression to Support Vector Machine, or any of the other learning algorithms found in scikit-learn, by changing exactly one line of code.
But do not be fooled. This is a powerful tool and using it involves a somewhat steep learning curve to use it properly and safely.
Best regards, Howard