NEVER grow a DataFrame!

TLDR; (just read the bold text)

Most answers here will tell you how to create an empty DataFrame and fill it out, but no one will tell you that it is a bad thing to do.

Here is my advice: Accumulate data in a list, not a DataFrame.

Use a list to collect your data, then initialise a DataFrame when you are ready. Either a list-of-lists or list-of-dicts format will work, pd.DataFrame accepts both.

data = []
for a, b, c in some_function_that_yields_data():
    data.append([a, b, c])

df = pd.DataFrame(data, columns=['A', 'B', 'C'])

Pros of this approach:

1. It is always cheaper to append to a list and create a DataFrame in one go than it is to create an empty DataFrame (or one of NaNs) and append to it over and over again.

2. Lists also take up less memory and are a much lighter data structure to work with, append, and remove (if needed).

3. dtypes are automatically inferred (rather than assigning object to all of them).

4. A RangeIndex is automatically created for your data, instead of you having to take care to assign the correct index to the row you are appending at each iteration.

If you aren't convinced yet, this is also mentioned in the documentation:

Iteratively appending rows to a DataFrame can be more computationally intensive than a single concatenate. A better solution is to append those rows to a list and then concatenate the list with the original DataFrame all at once.

But what if my function returns smaller DataFrames that I need to combine into one large DataFrame?

That's fine, you can still do this in linear time by growing or creating a python list of smaller DataFrames, then calling pd.concat.

small_dfs = []
for small_df in some_function_that_yields_dataframes():
    small_dfs.append(small_df)

large_df = pd.concat(small_dfs, ignore_index=True)

or, more concisely:

large_df = pd.concat(
    list(some_function_that_yields_dataframes()), ignore_index=True)

These options are horrible

append or concat inside a loop

Here is the biggest mistake I've seen from beginners:

df = pd.DataFrame(columns=['A', 'B', 'C'])
for a, b, c in some_function_that_yields_data():
    df = df.append({'A': i, 'B': b, 'C': c}, ignore_index=True) # yuck
    # or similarly,
    # df = pd.concat([df, pd.Series({'A': i, 'B': b, 'C': c})], ignore_index=True)

Memory is re-allocated for every append or concat operation you have. Couple this with a loop and you have a quadratic complexity operation.

The other mistake associated with df.append is that users tend to forget append is not an in-place function, so the result must be assigned back. You also have to worry about the dtypes:

df = pd.DataFrame(columns=['A', 'B', 'C'])
df = df.append({'A': 1, 'B': 12.3, 'C': 'xyz'}, ignore_index=True)

df.dtypes
A     object   # yuck!
B    float64
C     object
dtype: object

Dealing with object columns is never a good thing, because pandas cannot vectorize operations on those columns. You will need to do this to fix it:

df.infer_objects().dtypes
A      int64
B    float64
C     object
dtype: object

loc inside a loop

I have also seen loc used to append to a DataFrame that was created empty:

df = pd.DataFrame(columns=['A', 'B', 'C'])
for a, b, c in some_function_that_yields_data():
    df.loc[len(df)] = [a, b, c]

As before, you have not pre-allocated the amount of memory you need each time, so the memory is re-grown each time you create a new row. It's just as bad as append, and even more ugly.

Empty DataFrame of NaNs

And then, there's creating a DataFrame of NaNs, and all the caveats associated therewith.

df = pd.DataFrame(columns=['A', 'B', 'C'], index=range(5))
df
     A    B    C
0  NaN  NaN  NaN
1  NaN  NaN  NaN
2  NaN  NaN  NaN
3  NaN  NaN  NaN
4  NaN  NaN  NaN

It creates a DataFrame of object columns, like the others.

df.dtypes
A    object  # you DON'T want this
B    object
C    object
dtype: object

Appending still has all the issues as the methods above.

for i, (a, b, c) in enumerate(some_function_that_yields_data()):
    df.iloc[i] = [a, b, c]

The Proof is in the Pudding

Timing these methods is the fastest way to see just how much they differ in terms of their memory and utility.

Benchmarking code for reference.