Beancount-import is a tool for semi-automatically importing financial data from external data sources into the Beancount bookkeeping system, as well as merging and reconciling imported transactions with each other and with existing transactions.

Key features

• Pluggable data source architecture, including existing support for OFX (cash, investment, and retirement accounts), Mint.com, Amazon.com, and Venmo.

• Robustly associates imported transactions with the source data, to automatically avoid duplicates.

• Automatically predicts unknown legs of imported transactions based on a learned classifier (currently decision tree-based).

• Sophisticated transaction matching/merging system that can semi-automatically combine and reconcile both manually entered and imported transactions from independent sources.

• Easy-to-use, powerful web-based user interface.

Basic operation

From the data source modules, beancount-import obtains a list of pending imported transactions. (Balance and price entries may also be provided.) Depending on the external data source, pending transactions may fully specify all of the Beancount accounts (e.g. an investment transaction from an OFX source where shares of a stock are bought using cash in the same investment account), or may have some postings to unknown accounts, indicated by the special account name Expenses:FIXME. For example, pending transactions obtained from bank account/credit card account data (e.g. using the Mint.com data source) always have exactly two postings, one to the known Beancount account corresponding to the bank account from which the data was obtained, and the other to an unknown account.

For each pending transaction, beancount-import attempts to find matches to both existing transactions and to other pending transactions, and computes a set of candidate merged transactions. For each unknown account posting, Beancount-import predicts the account based on a learned classifier. Through a web interface, the user can view the pending transactions, select the original transaction or one of the merged candidates, and confirm or modify any predicted accounts. The web interface shows the lines in the journal that would be added or removed for each candidate. Once the user accepts a candidate, the candidate is inserted or merged into the Beancount journal, and the user is then presented with the next pending entry.

The imported transactions include metadata fields on the transaction and on the postings that serve several purposes:

• indicating to the data source module which entries in its external representation have already been imported and should not be imported again;
• indicating which postings are cleared, meaning they have been confirmed by the authoritative source, which also constrains matching (a cleared posting can only match an uncleared posting);
• providing necessary information for training the classifier used for predicting unknown accounts;
• providing information to the user that may be helpful for identifying and understanding the transaction.

Installation

1. Ensure you have activated a suitable Python 3 virtualenv if desired.

2. To install the most recent published package from PyPi, simply type:

   pip install beancount-import
   

   Alternatively, to install from a clone of the repository, type:

   python setup.py install
   

   or for development:

   python setup.py develop
   

Demo

To see Beancount-import in action on test data, refer to the instructions in the examples directory.

Data sources

Data sources are defined by implementing the Source interface defined by the beancount_import.source module.

The data sources provide a way to import and reconcile already-downloaded data. To retrieve financial data automatically, you can use the finance_dl package. You can also use any other mechanism, including manually downloading the data from a financial institution's website, provided that it is in the format required by the data source.

The currently supported set of data sources is:

• beancount_import.source.ofx: most versatile data source, supports checking, savings, credit card, investment, and retirement accounts.
• beancount_import.source.mint: supports all cash account types supported by Mint.com.
• beancount_import.source.venmo
• beancount_import.source.amazon: supports regular and digital order invoices.
• beancount_import.source.healthequity: supports HealthEquity.com HSA accounts, including both cash and investment transactions.
• beancount_import.source.google_purchases: imports purchases from that Google has heuristically extracted from Gmail messages.
• beancount_import.source.paypal: imports Paypal transactions.
• beancount_import.source.waveapps: imports receipts from Wave, a free receipt-scanning website/mobile app.

Refer to the individual data source documentation for details on configuration.

Usage

To run Beancount-import, create a Python script that invokes the beancount_import.webserver.main function. Refer to the examples fresh and manually_entered.

Errors

Any errors either from Beancount itself or one of the data sources are shown in the Errors tab. It is usually wise to manually resolve any errors, either using the built-in editor or an external editor, before proceeding, as some errors may result in incorrect behavior. Balance errors, however, are generally safe to ignore.

Viewing candidates

Select the Candidates tab to view the current pending imported entry, along with all proposed matches with existing and other pending transactions. The original unmatched entry is always listed last, and the proposal that includes the most matched postings is listed first. The list with checkboxes at the top indicates which existing or pending transactions are used in each proposed match; the current pending transaction is always listed first. If many incorrect matches were found, you can deselect the checkboxes to filter the matches.

You can select one of the proposed entries by clicking on it, or using the up/down arrow keys. To accept a proposed entry as is, you can press Enter or double click it. This immediately modifies the journal to reflect the change, and also displays the relevant portion of the journal in the Journal tab, so that you may easily make manual edits.

Specifying unknown accounts

If a proposed entry includes unknown accounts, they are highlighted with a distinctive background color and labeled with a group number. The account shown is the one that was automatically predicted, or Expenses:FIXME if automatic prediction was not possible (e.g. because of lack of training data). There are several ways to correct any incorrectly-predicted accounts:

• To change an individual account, you can Shift+click on it, type in the new account name, and then press Enter. If you press Escape while typing in the account name, the account will be left unchanged. A fuzzy matching algorithm is used for autocompletion: if you type "ex:co", for example, it will match any accounts for which there is a subsequence of 2 components, where the first starts (case-insensitively) with "ex" and the second starts with "co", such as an Expenses:Drinks:Coffee account.
• To change all accounts within a proposed entry that share the same group number, you can click on one of the accounts without holding shift, or press the digit key corresponding to the group number. Once you type in an account and press Enter, the specified account will be substituted for all postings in the group.
• To change all accounts within a proposed entry, you can click the Change account button or press the a key. Once you type in an account and press Enter, the specified account will be substituted for all unknown accounts in the current entry.
• If you wish to postpone specifying the correct account, you can click the Fixme later button or press the f key. This will substitute the original unknown account names for all unknown accounts in the current entry. If you then accept this entry, the transaction including these FIXME accounts will be added to your journal, and the next time you start Beancount-import the transaction will be treated as a pending entry.

Viewing associated source data

Data sources may indicate that additional source data is associated with particular candidate entries, typically based on the metadata fields and/or links that are included in the transaction. For example, the beancount_import.source.amazon data source associates the order invoice HTML page with the transaction, and the beancount_import.source.google_purchases data source associates the purchase details HTML page. Other possible source data types include PDF statements and receipt images.

You can view any associated source data for the currently selected candidate by selecting the Source data tab.

Changing the narration, payee, links or tags

To modify the narration of an entry, you can click on it, click the Narration button, or press the n key. This actually lets you modify the payee, links, and tags as well. If you introduce a syntax error in the first line of transaction, the text box will be highlighted in red and focus will remain until you either correct it or press Escape, which will revert the first line of the transaction back to its previous value.

Checking for uncleared postings

The Uncleared tab displays the list of postings to accounts for which there is an authoritative source and which have not been cleared. Normally, postings are marked as cleared by adding the appropriate source-specific metadata fields that associate it with the external data representation, such as an ofx_fitid field in the case of the OFX source.

This list may be useful for finding discrepancies that need manual correction. Typical causes of uncleared postings include:

1. The source data for the posting has not yet been downloaded.
2. The transaction is a duplicate of another transaction already in the journal, and needs to be manually merged/deleted.
3. The posting is from before the earliest date for which source data was imported, and no earlier data is available. Such postings can be ignored by adding a cleared_before: <date> metadata field to the open directive for the account or one of its ancestor accounts.
4. The source data is missing or cannot be imported, but the posting was manually verified. Such postings can be ignored by adding a cleared: TRUE metadata field to them.

Skipping and ignoring imported entries

If you are presented with a pending entry that you don't wish to import, you have several options:

1. You can skip past it by selecting a different transaction in the Pending tab, or can skip to the next pending entry by clicking on the button labeled ⏩ or pressing the ] key. This skips it in the current session, but it remains as a pending entry and will be included again if you restart beancount-import.

2. You can click on the button labeled Fixme later or press the f key to reset all unknown accounts, and then accept the candidate. This will add the transaction to your journal, but with the unknown accounts left as Expenses:FIXME. This is useful for transactions for which you don't know how to assign an account, or which you expect to match to another transaction that will be generated from data that hasn't yet been downloaded. Any transactions in the journal with Expenses:FIXME accounts will be included at the end of the list of pending entries the next time you start beancount-import.

3. You can click on the button labeled Ignore or press the i key to add the selected candidate to the special "ignored" journal file. This is useful for transactions that are erroneous, such as actual duplicates. Entries that are ignored will not be presented again if you restart beancount-import. However, if you manually delete them from the "ignored" journal file, they will return as pending entries.

Use with an existing Beancount journal

If you start using Beancount-import with an existing beancount journal containing transactions that are also referenced in the external data supplied to a data sources, the data source will not know to skip those transactions, because they will not have the requisite metadata indicating the association. Therefore, they will all be presented to you as new pending imported transactions.

However, the matching mechanism will very likely have determined the correct match to an existing transaction, which will be presented as the default option. Accepting these matches will simply have the effect of inserting the relevant metadata into your journal so that the transactions are considered "cleared" and won't be imported again next time you run Beancount-import. It should be a relatively quick process to do this even for a large number of transactions.

Development

For development of this package, make sure to install Beancount-import using the python3 setup.py develop command rather than python3 setup.py install. If you previously ran the install command, you can simply re-run setup.py with the develop command.

Testing

You can run the tests using the pytest command.

Many of the tests are "golden" tests, which work by creating a textual representation of some state and comparing it with the contents of a particular file in the testdata/ directory. If you change one of these tests or add a new one, you can have the tests automatically generate the output by setting the environment variable BEANCOUNT_IMPORT_GENERATE_GOLDEN_TESTDATA=1, e.g.:

BEANCOUNT_IMPORT_GENERATE_GOLDEN_TESTDATA=1 pytest

Make sure to commit to at least stage any changes you've made to the relevant testdata files prior to running the tests with this environment variable set. That way you can manually verify any changes between the existing output and the new output using git diff.

Web frontend

The web frontend source code is in the frontend/ directory. Refer to the README.md file there for how to rebuild and run the frontend after making changes.

Basic workflow

Simple expense transaction from Mint.com data source

Suppose the user has purchased a coffee at Starbucks on 2016-08-09 using a credit card, and has set up Mint.com to retrieve the transaction data for this credit card.

Given the following CSV entry:

"Date","Description","Original Description","Amount","Transaction Type","Category","Account Name","Labels","Notes"
"8/10/2016","Starbucks","STARBUCKS STORE 12345","2.45","debit","Coffee Shops","My Credit Card","",""

and the following open account directive:

1900-01-01 open Liabilities:Credit-Card  USD
  mint_id: "My Credit Card"

the Mint data source will generate the following pending transaction:

2016-08-10 * "STARBUCKS STORE 12345"
  Liabilities:Credit-Card             -2.45 USD
    date: 2016-08-10
    source_desc: "STARBUCKS STORE 12345"
  Expenses:FIXME                       2.45 USD

The user might manually specify that the unknown account is Expenses:Coffee. The web interface will then show the updated changeset:

+2016-08-10 * "STARBUCKS STORE 12345"
+  Liabilities:Credit-Card             -2.45 USD
+    date: 2016-08-10
+    source_desc: "STARBUCKS STORE 12345"
+  Expenses:Coffee                      2.45 USD

If the Expenses:Coffee account does not already exist, Beancount-import will additionally include an open directive in the changeset:

+2016-08-10 * "STARBUCKS STORE 12345"
+  Liabilities:Credit-Card             -2.45 USD
+    date: 2016-08-10
+    source_desc: "STARBUCKS STORE 12345"
+  Expenses:Coffee                      2.45 USD
+ 2016-08-10 open Expenses:Coffee USD

Once the user accepts this change, the changeset is applied to the journal. The presence of the date and source_desc metadata fields indicate to the Mint data source that the Liabilities:Credit-Card posting is cleared. The combination of the words in the source_desc, the source account of Liabilities:Credit-Card, and the target account of Expenses:Coffee serves as a training example for the classifier. A subsequent pending transaction with a source_desc field containing the word STARBUCKS is likely to be automatically classified as Expenses:Coffee. Note that while in this case the narration matches the source_desc field, the narration has no effect on the automatic prediction. The user must not delete or modify these metadata fields, but additional metadata fields may be added.

Mint.com has its own heuristics for computing the Description and Category fields from the Original Description provided by the financial institution. However, these are ignored by the Mint data source as they are not stable (can change if the data is re-downloaded) and not particularly reliable.

Match to a manually entered transaction

Considering the same transaction as shown in the previous example, suppose the user has already manually entered the transaction prior to running the import:

2016-08-09 * "Coffee"
  Liabilities:Credit-Card             -2.45 USD
  Expenses:Coffee

When running Beancount-import, the user will be presented with two candidates:

 2016-08-09 * "Coffee"
   Liabilities:Credit-Card             -2.45 USD
+    date: 2016-08-10
+    source_desc: "STARBUCKS STORE 12345"
   Expenses:Coffee


+2016-08-10 * "STARBUCKS STORE 12345"
+  Liabilities:Credit-Card             -2.45 USD
+    date: 2016-08-10
+    source_desc: "STARBUCKS STORE 12345"
+  Expenses:FIXME                       2.45 USD

The user should select the first one; selecting the second one would yield a duplicate transaction (but which could later be diagnosed as an uncleared transaction). The Expenses:FIXME account in the second candidate would in general actually be some other, possibly incorrect, predicted account, but which is clearly indicated as an prediction that can be changed.

As is typically the case, the date on the manually entered transaction (likely the date on which the transaction actually occurred) is not exactly the same as the date provided by the bank. To handle this discrepancy, Beancount-import allows matches between postings that are up to 5 days apart. The date metadata field allows the posting to be reliably matched to the corresponding entry in the CSV file, even though the overall transaction date differs.

Note that even though this transaction was manually entered, once it is matched with the pending transaction and the source_desc and date metadata fields are added, it functions as a training example exactly the same as in the previous example.

Credit card payment transaction

Suppose the user pays the balance of a credit card using a bank account, and Mint.com is set up to retrieve the transactions from both the bank account and the credit card.

Given the following CSV entries:

"Date","Description","Original Description","Amount","Transaction Type","Category","Account Name","Labels","Notes"
"11/27/2013","Transfer from My Checking","CR CARD PAYMENT ALEXANDRIA VA","66.88","credit","Credit Card Payment","My Credit Card","",""
"12/02/2013","National Federal Des","NATIONAL FEDERAL DES:TRNSFR","66.88","debit","Transfer","My Checking","",""

and the following open account directives:

1900-01-01 open Liabilities:Credit-Card  USD
  mint_id: "My Credit Card"

1900-01-01 open Assets:Checking  USD
  mint_id: "My Checking"

the Mint data source will generate 2 pending transactions, and for the first one will present two candidates:

+2013-11-27 * "CR CARD PAYMENT ALEXANDRIA VA"
+  Liabilities:Credit-Card             66.88 USD
+    date: 2013-11-27
+    source_desc: "CR CARD PAYMENT ALEXANDRIA VA"
+  Assets:Checking                    -66.88 USD
+    date: 2013-12-02
+    source_desc: "NATIONAL FEDERAL DES:TRNSFR"


+2013-11-27 * "CR CARD PAYMENT ALEXANDRIA VA"
+  Liabilities:Credit-Card             66.88 USD
+    date: 2013-11-27
+    source_desc: "CR CARD PAYMENT ALEXANDRIA VA"
+  Expenses:FIXME                     -66.88 USD

Note that the Expenses:FIXME account in the second transaction will actually be whichever account was predicted automatically. If there have been prior similar transactions, it is likely to be correct predicted as Assets:Checking.

The user should accept the first candidate to import both transactions at once. In that case, both postings are considered cleared, and the new transaction will result in two training examples for automatic prediction, corresponding to each of the two combinations of source_desc, source account, and target account.

However, if the user accepts the second candidate (perhaps because the transaction hasn't yet been posted to the checking account and the pending transaction derived from the checking account data is not yet available), and either leaves the account as Expenses:FIXME, manually specifies Assets:Checking, or relies on the automatic prediction to choose Assets:Checking, then when importing the transaction from the checking account, the user will be presented with the following candidates and will have another chance to accept the match:

 2013-11-27 * "CR CARD PAYMENT ALEXANDRIA VA"
   Liabilities:Credit-Card             66.88 USD
     date: 2013-11-27
     source_desc: "CR CARD PAYMENT ALEXANDRIA VA"
   Assets:Checking                    -66.88 USD
+    date: 2013-12-02
+    source_desc: "NATIONAL FEDERAL DES:TRNSFR"


+2013-12-02 * "NATIONAL FEDERAL DES:TRNSFR"
+  Assets:Checking                    -66.88 USD
+    date: 2013-12-02
+    source_desc: "NATIONAL FEDERAL DES:TRNSFR"
+  Expenses:FIXME                      66.88 USD

License

Copyright (C) 2014-2018 Jeremy Maitin-Shepard.

Distributed under the GNU General Public License, Version 2.0 only. See LICENSE file for details.