conservancy_beancount/conservancy_beancount/reconcile/statement_reconciler.py

631 lines
26 KiB
Python

"""Compare a bank CSV statement with the books.
This tool takes an AMEX or First Republic CSV statement file and
compares it line-by-line with the Beancount books to make sure that
everything matches. This is designed for situations where transactions
are entered into the books directly, rather than being imported from a
statement after the fact.
The reconciler will attempt to match transactions based on date,
amount, check number and payee, but is forgiving to differences in
dates, the absensce of check number and inexact matches on
payee. Matches are ranked, so where there is only one decent match for
an amount/date this is accepted, but if there are multiple similar
candidates it will refuse to guess.
The reconciler will also attempt to identify where a single statement
entry has been split out into multiple Beancount postings, such as a
single bank transfer representing health insurance for multiple
employees.
Run it like this:
$ statement_reconciler \
--beancount-file 2021.beancount \
--account Liabilities:CreditCard:AMEX \
--csv-statement ~/svn/2021-09-10_AMEX_activity.csv \
--bank-statement ~/svn/2021-09-10_AMEX_activity.pdf
Background:
Beancount users often write importers to create bookkeeping entries
direct from a bank statement or similar. That approach automates data
entry and reconciliation in one step. In some cases though, it's
useful to manually enter transactions and reconcile them later
on. This workflow helpful in cases like writing a paper check when
there's a time lag between committing to making a payment and the
funds being debited. That's the workflow we're using here.
Conservancy currently enter data by hand rather than using Beancount
importers. This tool is still somewhat like an importer in that it
needs to extract transaction details from a third-party
statement. Instead of creating directives, it just checks to see that
similar directives are already present. This is a bit like diff-ing a
statement with the books (though we're only interested in the presence
of lines, not so much their order).
Problems in scope:
- errors in the books take hours to find during reconciliation,
requiring manually comparing statemnts and the books and are
succeptible to mistakes, such as not noticing when there are two
payments for the same amount on the statement, but not in the books
("you're entering a world of pain")
- adding statement/reconciliation metadata to books is/was manual and
prone to mistakes
- Beancount doesn't provide any infrastructure for programmatically
updating the books, only appending in the case of importers
- paper checks are entered in the books when written, but may not be
cashed until months later (reconcile errors)
- jumping to an individual transaction in a large ledger isn't
trivial - Emacs grep mode is the current best option
- Pam and other staff don't use Emacs
- auditors would prefer Bradley didn't perform reconciliation,
ideally not Rosanne either
- reconciliation reports are created by hand when there are mismatches
Other related problems we're not dealing with here:
- after updates to the books files, beancount must be restarted to
reflect updates
- updates also invalidate the cache meaning restart takes several
minutes
- balance checks are manually updated in
svn/Financial/Ledger/sanity-check-balances.yaml
- transactions are entered manually and reconciled after the fact,
but importing from statements may be useful in some cases
"""
# TODO:
# - extract the magic numbers
# - consider merging in helper.py
import argparse
import collections
import copy
import csv
import datetime
import decimal
import io
import itertools
import logging
import os
import re
import sys
from typing import Callable, Dict, List, Optional, Sequence, Tuple, TextIO
from beancount import loader
from beancount.query.query import run_query
from colorama import Fore, Style # type: ignore
from .. import cliutil
from .. import config as configmod
if not sys.warnoptions:
import warnings
# Disable annoying warning from thefuzz prompting for a C extension. The
# current pure-Python implementation isn't a bottleneck for us.
warnings.filterwarnings('ignore', category=UserWarning, module='thefuzz.fuzz')
from thefuzz import fuzz # type: ignore
PROGNAME = 'reconcile-statement'
logger = logging.getLogger(__name__)
# Get some interesting feedback on call to RT with this:
# logger.setLevel(logging.DEBUG)
# logger.addHandler(logging.StreamHandler())
JUNK_WORDS = [
'software',
'freedom',
'conservancy',
'conse',
'payment',
'echeck',
'bill',
'debit',
'wire',
'credit',
"int'l",
"in.l",
'llc',
'online',
'donation',
'usd',
'inc',
]
JUNK_WORDS_RES = [re.compile(word, re.IGNORECASE) for word in JUNK_WORDS]
ZERO_RE = re.compile('^0+')
def remove_duplicate_words(text: str) -> str:
unique_words = []
known_words = set()
for word in text.split():
if word.lower() not in known_words:
unique_words.append(word)
known_words.add(word.lower())
return ' '.join(unique_words)
def remove_payee_junk(payee: str) -> str:
"""Clean up payee field to improve quality of fuzzy matching.
It turns out that bank statement "description" fields are
difficult to fuzzy match on because they're long and
noisey. Truncating them (see standardize_XXX_record fns) and
removing the common junk helps significantly.
"""
for r in JUNK_WORDS_RES:
payee = r.sub('', payee)
payee = ZERO_RE.sub('', payee)
payee = payee.replace(' - ', ' ')
payee = re.sub(r'\.0\.\d+', ' ', payee)
payee = payee.replace('.0', ' ')
payee = payee.replace('/', ' ')
payee = re.sub(re.escape('.com'), ' ', payee, flags=re.IGNORECASE)
payee = re.sub(re.escape('.net'), ' ', payee, flags=re.IGNORECASE)
payee = payee.replace('*', ' ')
payee = ' '.join([i for i in payee.split(' ') if len(i) > 2])
payee = payee.replace('-', ' ')
payee = remove_duplicate_words(payee)
payee.strip()
return payee
def read_transactions_from_csv(f: TextIO, standardize_statement_record: Callable) -> list:
reader = csv.DictReader(f)
# The reader.line_num is the source line number, not the spreadsheet row
# number due to multi-line records.
return sort_records([standardize_statement_record(row, i) for i, row in enumerate(reader, 2)])
def validate_amex_csv(sample: str, account: str) -> None:
required_cols = {'Date', 'Amount', 'Description', 'Card Member'}
reader = csv.DictReader(io.StringIO(sample))
if reader.fieldnames and not required_cols.issubset(reader.fieldnames):
sys.exit(f"This CSV doesn't seem to have the columns we're expecting, including: {', '.join(required_cols)}")
def standardize_amex_record(row: Dict, line: int) -> Dict:
"""Turn an AMEX CSV row into a standard dict format representing a transaction."""
# NOTE: Statement doesn't seem to give us a running balance or a final total.
return {
'date': datetime.datetime.strptime(row['Date'], '%m/%d/%Y').date(),
'amount': -1 * decimal.Decimal(row['Amount']),
# Descriptions have too much noise, so taking just the start
# significantly assists the fuzzy matching.
'payee': remove_payee_junk(row['Description'] or '')[:20],
'check_id': '',
'line': line,
}
def validate_fr_csv(sample: str, account: str) -> None:
required_cols = {'Date', 'Amount', 'Detail', 'Serial Num'}
reader = csv.DictReader(io.StringIO(sample))
if reader.fieldnames and not required_cols.issubset(reader.fieldnames):
sys.exit(f"This CSV doesn't seem to have the columns we're expecting, including: {', '.join(required_cols)}")
def standardize_fr_record(row: Dict, line: int) -> Dict:
return {
'date': datetime.datetime.strptime(row['Date'], '%m/%d/%Y').date(),
'amount': decimal.Decimal(row['Amount']),
'payee': remove_payee_junk(row['Detail'] or '')[:20],
'check_id': row['Serial Num'].lstrip('0'),
'line': line,
}
def standardize_beancount_record(row) -> Dict: # type: ignore[no-untyped-def]
"""Turn a Beancount query result row into a standard dict representing a transaction."""
return {
'date': row.date,
'amount': row.number_cost_position,
'payee': remove_payee_junk(f'{row.payee or ""} {row.entity or ""} {row.narration or ""}'),
'check_id': str(row.check_id or ''),
'filename': row.filename,
'line': row.line,
'bank_statement': row.bank_statement,
}
def format_record(record: dict) -> str:
if record['payee'] and record['check_id']:
output = f"{record['date'].isoformat()}: {record['amount']:11,.2f} {record['payee'][:25]} #{record['check_id']}".ljust(59)
elif record['payee']:
output = f"{record['date'].isoformat()}: {record['amount']:11,.2f} {record['payee'][:35]}".ljust(59)
else:
output = f"{record['date'].isoformat()}: {record['amount']:11,.2f} #{record['check_id']}".ljust(59)
return output
def format_multirecord(r1s: list[dict], r2s: list[dict], note: str) -> list[list]:
assert len(r1s) == 1
assert len(r2s) > 1
match_output = []
match_output.append([r1s[0]['date'], f'{format_record(r1s[0])}{format_record(r2s[0])} ✓ Matched{note}'])
for r2 in r2s[1:]:
match_output.append([r1s[0]['date'], f'{r1s[0]["date"].isoformat()}: ↳ → {format_record(r2)} ✓ Matched{note}'])
return match_output
def sort_records(records: List) -> List:
return sorted(records, key=lambda x: (x['date'], x['amount']))
def first_word_exact_match(a: str, b: str) -> float:
if len(a) == 0 or len(b) == 0:
return 0.0
first_a = a.split()[0].strip()
first_b = b.split()[0].strip()
if first_a.casefold() == first_b.casefold():
return min(1.0, 0.2 * len(first_a))
else:
return 0.0
def payee_match(a: str, b: str) -> float:
fuzzy_match = float(fuzz.token_set_ratio(a, b) / 100.00)
first_word_match = first_word_exact_match(a, b)
return max(fuzzy_match, first_word_match)
def records_match(r1: Dict, r2: Dict) -> Tuple[float, List[str]]:
"""Do these records represent the same transaction?"""
date_score = date_proximity(r1['date'], r2['date'])
if r1['date'] == r2['date']:
date_message = ''
elif date_score > 0.0:
diff = abs((r1['date'] - r2['date']).days)
date_message = f'+/- {diff} days'
else:
date_message = 'date mismatch'
if r1['amount'] == r2['amount']:
amount_score, amount_message = 2.0, ''
else:
amount_score, amount_message = 0.0, 'amount mismatch'
# We never consider payee if there's a check_id in the books.
check_message = ''
payee_message = ''
# Sometimes we get unrelated numbers in the statement column with check-ids,
# so we can't match based on the existence of a statement check-id.
if r2['check_id']:
payee_score = 0.0
if r1['check_id'] and r2['check_id'] and r1['check_id'] == r2['check_id']:
check_score = 1.0
else:
check_message = 'check-id mismatch'
check_score = 0.0
else:
check_score = 0.0
payee_score = payee_match(r1['payee'], r2['payee'])
if payee_score > 0.8:
payee_message = ''
elif payee_score > 0.4:
payee_message = 'partial payee match'
else:
payee_message = 'payee mismatch'
overall_score = (date_score + amount_score + check_score + payee_score) / 4
overall_message = [m for m in [date_message, amount_message, check_message, payee_message] if m]
return overall_score, overall_message
def match_statement_and_books(statement_trans: List[Dict], books_trans: List[Dict]) -> Tuple[List[Tuple[List, List, List]], List[Dict], List[Dict]]:
"""
Runs through all the statement transactions to find a matching transaction
in the books. If found, the books transaction is marked off so that it can
only be matched once. Some transactions will be matched, some will be on the
statement but not the books and some on the books but not the statement.
"""
matches = []
remaining_books_trans = []
remaining_statement_trans = []
for r1 in statement_trans:
best_match_score = 0.0
best_match_index = None
best_match_note = []
matches_found = 0
for i, r2 in enumerate(books_trans):
score, note = records_match(r1, r2)
if score >= 0.5 and score >= best_match_score:
matches_found += 1
best_match_score = score
best_match_index = i
best_match_note = note
if best_match_score > 0.5 and matches_found == 1 and 'check-id mismatch' not in best_match_note or best_match_score > 0.8:
matches.append(([r1], [books_trans[best_match_index]], best_match_note))
# Don't try to make a second match against this books entry.
if best_match_index is not None:
del books_trans[best_match_index]
else:
remaining_statement_trans.append(r1)
for r2 in books_trans:
remaining_books_trans.append(r2)
return matches, remaining_statement_trans, remaining_books_trans
# TODO: Return list of tuples (instead of list of lists).
def format_matches(matches: List, csv_statement: str, show_reconciled_matches: bool) -> List[List]:
match_output = []
for r1s, r2s, note in matches:
note = ', '.join(note)
note = ': ' + note if note else note
if r1s and r2s:
if show_reconciled_matches or not all(x['bank_statement'] for x in r2s):
if len(r2s) == 1:
entry = [r1s[0]['date'], f'{format_record(r1s[0])}{format_record(r2s[0])} ✓ Matched{note}']
if 'payee mismatch' in note:
entry[1] = Fore.YELLOW + Style.BRIGHT + entry[1] + Style.RESET_ALL
match_output.append(entry)
else:
match_output.extend(format_multirecord(r1s, r2s, note))
elif r1s:
match_output.append([r1s[0]['date'], Fore.RED + Style.BRIGHT + f'{format_record(r1s[0])}{" ":^59} ✗ NOT IN BOOKS ({os.path.basename(csv_statement)}:{r1s[0]["line"]})' + Style.RESET_ALL])
else:
match_output.append([r2s[0]['date'], Fore.RED + Style.BRIGHT + f'{" ":^59}{format_record(r2s[0])} ✗ NOT ON STATEMENT ({os.path.basename(r2s[0]["filename"])}:{r2s[0]["line"]})' + Style.RESET_ALL])
return match_output
def date_proximity(d1: datetime.date, d2: datetime.date) -> float:
diff = abs(int((d1 - d2).days))
if diff > 60:
return 0.0
else:
return 1.0 - (diff / 60.0)
def metadata_for_match(match: Tuple[List, List, List], statement_filename: str, csv_filename: str) -> List[Tuple[str, int, str]]:
# Can we really ever have multiple statement entries? Probably not.
statement_filename = get_repo_relative_path(statement_filename)
csv_filename = get_repo_relative_path(csv_filename)
metadata = []
statement_entries, books_entries, _ = match
for books_entry in books_entries:
for statement_entry in statement_entries:
if not books_entry['bank_statement']:
metadata.append((books_entry['filename'], books_entry['line'], f' bank-statement: "{statement_filename}"'))
metadata.append((books_entry['filename'], books_entry['line'], f' bank-statement-csv: "{csv_filename}:{statement_entry["line"]}"'))
return metadata
def write_metadata_to_books(metadata_to_apply: List[Tuple[str, int, str]]) -> None:
"""Insert reconciliation metadata in the books files.
Takes a list of edits to make as tuples of form (filename, lineno, metadata):
[
('2021/main.beancount', 4245, ' bank-statement: statement.pdf'),
('2021/main.beancount', 1057, ' bank-statement: statement.pdf'),
('2021/payroll.beancount', 257, ' bank-statement: statement.pdf'),
...,
]
"""
file_contents: dict[str, list] = {}
file_offsets: dict[str, int] = collections.defaultdict(int)
# Load each books file into memory and insert the relevant metadata lines.
# Line numbers change as we do this, so we keep track of the offset for each
# file. Changes must be sorted by line number first or else the offsets will
# break because we're jumping around making edits.
for filename, line, metadata in sorted(metadata_to_apply):
if filename not in file_contents:
with open(filename, 'r') as f:
file_contents[filename] = f.readlines()
# Insert is inefficient, but fast enough for now in practise.
file_contents[filename].insert(line + file_offsets[filename], metadata.rstrip() + '\n')
file_offsets[filename] += 1
# Writes each updated file back to disk.
for filename, contents in file_contents.items():
with open(filename, 'w') as f:
f.writelines(contents)
print(f'Wrote {filename}.')
def get_repo_relative_path(path: str) -> str:
return os.path.relpath(path, start=os.getenv('CONSERVANCY_REPOSITORY'))
def parse_path(path: str) -> str:
if not os.path.exists(path):
raise argparse.ArgumentTypeError(f'File {path} does not exist.')
return path
def parse_repo_relative_path(path: str) -> str:
if not os.path.exists(path):
raise argparse.ArgumentTypeError(f'File {path} does not exist.')
repo = os.getenv('CONSERVANCY_REPOSITORY')
if not repo:
raise argparse.ArgumentTypeError('$CONSERVANCY_REPOSITORY is not set.')
if not path.startswith(repo):
raise argparse.ArgumentTypeError(f'File {path} does not share a common prefix with $CONSERVANCY_REPOSITORY {repo}.')
return path
def parse_decimal_with_separator(number_text: str) -> decimal.Decimal:
"""decimal.Decimal can't parse numbers with thousands separator."""
number_text = number_text.replace(',', '')
return decimal.Decimal(number_text)
def parse_arguments(argv: List[str]) -> argparse.Namespace:
parser = argparse.ArgumentParser(prog=PROGNAME, description='Reconciliation helper')
cliutil.add_version_argument(parser)
cliutil.add_loglevel_argument(parser)
parser.add_argument('--beancount-file', required=True, type=parse_path)
parser.add_argument('--csv-statement', required=True, type=parse_repo_relative_path)
parser.add_argument('--bank-statement', required=True, type=parse_repo_relative_path)
parser.add_argument('--account', required=True, help='eg. Liabilities:CreditCard:AMEX')
# parser.add_argument('--report-group-regex')
parser.add_argument('--show-reconciled-matches', action='store_true')
parser.add_argument('--non-interactive', action='store_true', help="Don't prompt to write to the books") # parser.add_argument('--statement-balance', type=parse_decimal_with_separator, required=True, help="A.K.A \"cleared balance\" taken from the end of the period on the PDF statement. Required because CSV statements don't include final or running totals")
args = parser.parse_args(args=argv)
return args
def totals(matches: List[Tuple[List, List, List]]) -> Tuple[decimal.Decimal, decimal.Decimal, decimal.Decimal]:
total_matched = decimal.Decimal(0)
total_missing_from_books = decimal.Decimal(0)
total_missing_from_statement = decimal.Decimal(0)
for statement_entries, books_entries, _ in matches:
if statement_entries and books_entries:
total_matched += sum(c['amount'] for c in statement_entries)
elif statement_entries:
total_missing_from_books += sum(c['amount'] for c in statement_entries)
else:
total_missing_from_statement += sum(c['amount'] for c in books_entries)
return total_matched, total_missing_from_books, total_missing_from_statement
def subset_match(statement_trans: List[dict], books_trans: List[dict]) -> Tuple[List[Tuple[List, List, List]], List[Dict], List[Dict]]:
matches = []
remaining_books_trans = []
remaining_statement_trans = []
groups = itertools.groupby(books_trans, key=lambda x: (x['date'], x['payee']))
for _, group in groups:
best_match_score = 0.0
best_match_index = None
best_match_note = []
matches_found = 0
group_items = list(group)
total = sum(x['amount'] for x in group_items)
r2 = copy.copy(group_items[0])
r2['amount'] = total
for i, r1 in enumerate(statement_trans):
score, note = records_match(r1, r2)
if score >= 0.5 and score >= best_match_score:
matches_found += 1
best_match_score = score
best_match_index = i
best_match_note = note
if best_match_score > 0.5 and matches_found == 1 and 'check-id mismatch' not in best_match_note or best_match_score > 0.8:
matches.append(([statement_trans[best_match_index]], group_items, best_match_note))
if best_match_index is not None:
del statement_trans[best_match_index]
else:
remaining_books_trans.append(r2)
for r1 in statement_trans:
remaining_statement_trans.append(r1)
return matches, remaining_statement_trans, remaining_books_trans
def process_unmatched(statement_trans: List[dict], books_trans: List[dict]) -> List[Tuple[List, List, List]]:
matches: List[Tuple[List, List, List]] = []
for r1 in statement_trans:
matches.append(([r1], [], ['no match']))
for r2 in books_trans:
matches.append(([], [r2], ['no match']))
return matches
def main(arglist: Optional[Sequence[str]] = None,
stdout: TextIO = sys.stdout,
stderr: TextIO = sys.stderr,
config: Optional[configmod.Config] = None,
) -> int:
args = parse_arguments(arglist)
cliutil.set_loglevel(logger, args.loglevel)
if config is None:
config = configmod.Config()
config.load_file()
# TODO: Should put in a sanity check to make sure the statement you're feeding
# in matches the account you've provided.
# TODO: Can we open the files first, then pass the streams on to the rest of the program?
if 'AMEX' in args.account:
validate_csv = validate_amex_csv
standardize_statement_record = standardize_amex_record
else:
validate_csv = validate_fr_csv
standardize_statement_record = standardize_fr_record
with open(args.csv_statement) as f:
sample = f.read(200)
validate_csv(sample, args.account)
f.seek(0)
statement_trans = read_transactions_from_csv(f, standardize_statement_record)
begin_date = statement_trans[0]['date']
end_date = statement_trans[-1]['date']
# Do we traverse and filter the in-memory entries list and filter that, or do we
# use Beancount Query Language (BQL) to get a list of transactions? Currently
# using BQL.
#
# beancount.query.query_compile.compile() and
# beancount.query.query_execute.filter_entries() look useful in this respect,
# but I'm not clear on how to use compile(). An example would help.
entries, _, options = loader.load_file(args.beancount_file)
# books_balance_query = f"""SELECT sum(COST(position)) AS aa WHERE account = "{args.account}"
# AND date <= {end_date.isoformat()}"""
# _, result_rows = run_query(entries, options, books_balance_query, numberify=True)
# books_balance = result_rows[0][0] if result_rows else 0
# String concatenation looks bad, but there's no SQL injection possible here
# because BQL can't write back to the Beancount files. I hope!
query = f'SELECT filename, META("lineno") AS line, META("bank-statement") AS bank_statement, date, number(cost(position)), payee, ENTRY_META("entity") as entity, ANY_META("check-id") as check_id, narration where account = "{args.account}" and date >= {begin_date} and date <= {end_date}'
_, result_rows = run_query(entries, options, query)
books_trans = sort_records([standardize_beancount_record(row) for row in result_rows])
matches, remaining_statement_trans, remaining_books_trans = match_statement_and_books(statement_trans, books_trans)
subset_matches, remaining_statement_trans, remaining_books_trans = subset_match(remaining_statement_trans, remaining_books_trans)
matches.extend(subset_matches)
unmatched = process_unmatched(remaining_statement_trans, remaining_books_trans)
matches.extend(unmatched)
match_output = format_matches(matches, args.csv_statement, args.show_reconciled_matches)
_, total_missing_from_books, total_missing_from_statement = totals(matches)
print('-' * 155)
statement_heading = f'Statement transactions {begin_date} to {end_date}'
print(f'{statement_heading:<52} {"Books transactions":<58} Notes')
print('-' * 155)
for _, output in sorted(match_output, key=lambda x: x[0]):
print(output)
print('-' * 155)
print(f'Sub-total not on statement: {total_missing_from_statement:12,.2f}')
print(f'Sub-total not in books: {total_missing_from_books:12,.2f}')
print(f'Total: {total_missing_from_statement + total_missing_from_books:12,.2f}')
print('-' * 155)
# Write statement metadata back to books
metadata_to_apply = []
for match in matches:
metadata_to_apply.extend(metadata_for_match(match, args.bank_statement, args.csv_statement))
if metadata_to_apply and not args.non_interactive:
print('Mark matched transactions as reconciled in the books? (y/N) ', end='')
if input().lower() == 'y':
write_metadata_to_books(metadata_to_apply)
entry_point = cliutil.make_entry_point(__name__, PROGNAME)
if __name__ == '__main__':
exit(entry_point())