Writing Solid Code¶
Break down a process into chunks.
Package each chunk into a separate function.
A key idea: aside from an explicitly specified set of inputs and outputs, the function is independent of the rest of the world (encapsulation).
Write your code top down by stepwise refinement
- Start with an outline of the steps.
- For each step:
- if it’s trivial: write it out in pseudo code.
- it it’s not: write another outline for that step
- Finally, translate the pseudo code into code.
Each function should be short and perform exactly one task.
Write reusable code¶
- Write each function to be sufficiently general, so it can be reused in other projects.
- Over time, you will accumulate a library of code that can be used over and over.
- Example: Write code that produces marginal products, average costs, etc for a CES production function.
Notes on Writing Code¶
Read a good book on best practices in programming.
- I see a lot of very poorly written code that is impossible to understand and not robust. Do yourself a favor and save a lot of time down the road by learning to how write quality code.
- A book I like is “Writing Solid Code.”
- No literals as in
i1 = 1 : 57. It’s not robust and hard to read.
- No global variables.
- Don’t worry about speed. Worry about robustness and transparency.
- Unique names: I suffix all functions I write with a project code (e.g.
var_save_sc.m, etc). It avoids naming conflicts with other projects.
- Your code should contain lots of self-testing code. Most code is so fast that the loss of speed is irrelevant. If it is relevant, have a switch that globally switches test code on and off.
- Avoid using reserved words, in particular
ias an index.
This point is hard to overstate. It is extremely important to write code that is easy to understand and easy to maintain.
In practice, you often revisit programs months or years after they were written. They need to be well documented and well structured.
The programs needed to solve a stochastic OLG model have thousands of lines of code. The only way to understand something this complex is to break it into logical, self-contained pieces (a function that solves the household problem, another that solves the firm problem, etc.).
One example of how important this is:
Air traffic control centers still operate with hardware from the 1970s. The reason is that nobody understands the software well enough to port it to new hardware.
The FAA has already spent billions of dollars on unsuccessful attempts to rewrite this mess.
Another example is the Space Shuttle, which runs (now “ran”) on hardware from the 1960s. The reason is again that the software engineers can no longer understand the existing code.
There are many books on good programming style. One that I like is Writing Solid Code by Steve Maguire. Read it!
Your code should rarely use specific values for any object. When you refer to an object, do so by its name.
For example, create variables to hold directory names and constants. The reason is that code is otherwise hard to change and maintain.
Imagine you set some parameter
sigma=2, but refer to it as
2 in your code instead of
sigma. If you decide to try
sigma = 3, you need to locate and change every occurrence of
sigma in your code. It’s a mess.
The Golden Rule is: Every literal must have a name. Its value is defined in one place only.
Related to this: do not hard-code functional forms.
If you want to compute the marginal product of capital, write a function for it. Otherwise, if you want to switch from Cobb-Douglas to CES, you have to rewrite all your programs.
- Object oriented programming makes it easy to swap out entire parts of a model. We will talk about this later.
Your code should test itself automatically and periodically.
Embed error catching code everywhere (use valideattributes).
Catching bugs early makes them easier to find.
A trick to prevent your code from getting slowed down by self-testing:
- add a debugging switch as an input argument to each function (I call it
dbgis 0: go for speed and turn off self-testing
dbg > 10, run all self-test code
The process is then:
- Write code. Make sure it runs (correct syntax).
- Make sure it is correct (run all self-test code – slow)
- When you are confident that your code is good, set
dbg = 0and go for speed
- But every now and then, randomly switch
dbgon so that self tests are run (little cost in terms of run time; a lot of gain in terms of confidence in your code).
Automated Unit Testing¶
The golden rule:
When you write a function, write a test function to go with it.
It is hard to overstate the importance of automated testing. It gives you peace of mind. When you change some code, you can simply rerun your test suite and ensure that nothing has been broken.
The key is to fully automate the testing. Your project should have a single function that runs all tests in order.
All programming languages have unit testing frameworks that make it easy to automate this process. Matlab’s framework is described here.
Optimization refers to program modifications that speed up execution.
Think before you optimize!
Most code runs so fast that optimization is simply a waste of time.
Also: Beware of your intuition about where the program spends most of its time.
Here is an example: Consider the function that solves a stochastic OLG model.
It turns out that it spends 80% of its time running the Matlab interpolation function
There is little point optimizing the rest of the code.
To find out what makes your program slow, run the Matlab profiler.
Some of Matlab’s built-in functions are extremely slow.
- Two examples are
- It is easy to write replacements that run ten times faster.
- The Lightspeed library contains faster versions of built-in functions.
Passing arguments in the wrong order.¶
Matlab does not check the types of arguments.
Often functions have lots of input arguments.
It is easy to confuse the order and write
myfun(b,a) instead of
To avoid this: check that inputs have admissible values.
Passing too few arguments.¶
Matlab permits to omit input or output arguments when calling a function.
It is useful to check that the number of input arguments is as expected using nargin.
Reusing variable names.¶
Matlab does not permit explicit declaration of variables. It is therefore easy to use a variable name twice without noticing.
It is easy to make mistakes when extracting elements from matrices. This is especially true for code that wraps a loop into a single line of code.
For example, this is easy to read:
for ix = 1 : nx zV(ix) = xV(ix+2) + yV(nx + 2 - ix); end
This is the same thing, more compact but harder to read:
zV = xV(3 : nx+2) + yV(nx+1 : -1 : 2);
Tip: Write out code explicitly. Once it works, one can still make it faster (if that is even worthwhile).
Another common indexing mistake is to use too few arguments. For example:
x = rand([3,4]); y = x(3);
This should produce a syntax error, but it does not. Instead, it flattens x into a vector and then takes the 3rd element.
Material for Economists¶
Quantitative Economics by Sargent and Stachursky
- a really nice collection of lectures and exercises that covers both programming and the economics of the material (in Julia and Python)