One month · TD chequing · May 2026

A $3,950 paycheque on May 1st.
Overdrawn by May 31st.

Ryan D'Souza is a 24-year-old data analyst in Toronto, nine months into his first job. He wants a $3,000 emergency fund but keeps ending the month at zero. This shows where he is wasting money, the opportunity cost of that spending, and his plan to reach $3,000.

Balance across the month1
Income
Spent
Balance, May 31
Recoverable, no sacrifice
The finding

Follow the paycheque

How $3,950 becomes an overdraft

The month begins with a small balance and a full paycheque, and ends below zero. Each downward step is a group of spending. Amber marks the groups with the most waste.

What "waste" means here

Three tiers, not one number

Waste is not the same as spending. This separates money that bought nothing from money Ryan cannot see, and both of those from lifestyle he chose. Every figure opens to the transactions behind it.

How each transaction is sorted into a tier
The whole month

Where the money went

Every dollar of spending, grouped. Amber marks the categories with the most waste.

Daily spending across May

Most of the spending happens early in the month

Darker cells are heavier spending days. Most of the spending happens in the first half of the month, and the balance falls steadily until the overdraft on May 31.

How the categorization was checked

AI, cross-checked by rules

Claude, a frontier model from Anthropic, categorized every transaction. A separate rules engine did the same, independently. The score below is the model against a hand-labelled ground truth. The table is where the two methods disagreed, which is where a person should look.

The cross-check, step by step
Model accuracy

When AI and rules disagreed

transactions. The model was right on , the rule was right on . Neither is trustworthy on its own.

This is the whole reason for the cross-check. A single wrong category can flip Ryan's decision about what to cut, and frontier models do make errors on financial data.5

The disagreements
TransactionAI saidRules saidTruthRight
What the AI got wrong
    Flagged by the model for human review

    Low-confidence calls the model would not make on its own.

    The point of all this

    Ryan's path to $3,000

    Choose what Ryan changes and the timeline updates. The first three are the no-sacrifice cuts. The last two are his call.

    Time to $3,000
    months

    The opportunity cost of not saving it6

    Put in a 3% account, the same amount would be after one year, of which is interest Ryan gives up by spending it instead.

    The answer to his question

    What I would tell Ryan

    Plain steps, ordered by effort. The first four take an afternoon and cost nothing to do.

    1

    Cancel five unused subscriptions this week

    LinkedIn Premium, Disney+, Amazon Prime, the Netflix add-on, and Crave.2 He named Netflix and Spotify as the only two he uses, so the other five are unused. Crave was also billed twice this month, so cancelling it stops the fee and the duplicate is a separate one-time refund.

    $92/mo
    2

    Set a low-balance alert

    The overdraft fee happened because he did not see zero coming. A free TD alert defaulting to a $100 balance removes it.7

    $45/mo
    3

    Move the first $250 on payday, before spending

    Automate a transfer the day salary lands, ideally into a TFSA so the interest is tax-free.8, 9 He never sees it, so he cannot spend it.

    builds the fund
    4

    Use TD machines and track the cash

    Three withdrawals were at non-TD machines that add a surcharge.4 The bigger issue is that $560 in cash left with no record. A weekly cash cap makes it visible.

    ~$280/mo
    5

    Decide on the gym

    At one or two visits a month he is paying $30 to $60 a visit.3 Commit to using it, or switch to a pay-per-visit pass.

    $60/mo

    Steps 1, 2, and 4 alone recover about $417 a month with nothing Ryan enjoys removed, which reaches $3,000 in roughly seven months. A starter fund of this size is a sound first goal,10 and Ryan is not alone in struggling to save: the Canadian household saving rate was 4.4% late in 2025.11

    How this is built

    The pipeline

    The statement runs through two independent analyses and an evaluation, then into a static page that makes no live model call.

    References and sources 11 sources
    1. Ryan D'Souza's May 2026 TD chequing statement, the dataset analysed here (provided for the challenge).
    2. C+R Research, "Subscription Service Statistics and Costs" (consumers underestimate subscription spending by about $133 a month, and 42% have forgotten a paid subscription while still being charged). crresearch.com
    3. DellaVigna, S., and Malmendier, U. (2006). "Paying Not to Go to the Gym." American Economic Review 96(3): 694-719. aeaweb.org
    4. TD Canada Trust, "List of Services and Fees," effective March 5, 2026 (non-TD ATM withdrawal $2.00). td.com
    5. "Deficiency of Large Language Models in Finance: An Empirical Examination of Hallucination" (arXiv 2311.15548). arxiv.org
    6. Bank of Canada policy rate held at 2.25% (June 2026), and Ratehub high-interest savings rates of about 2 to 3%. bankofcanada.ca, ratehub.ca
    7. TD Canada Trust, "TD Alerts" (free threshold alerts, default at a $100 balance). td.com
    8. Thaler, R., and Benartzi, S. (2004). "Save More Tomorrow." Journal of Political Economy 112(S1): S164-S187. journals.uchicago.edu
    9. Canada Revenue Agency, TFSA contribution limits (2026 limit $7,000). canada.ca
    10. Financial Consumer Agency of Canada, "Setting up an emergency fund." canada.ca
    11. Statistics Canada, "Distributions of household economic accounts, Q4 2025" (household saving rate 4.4%). statcan.gc.ca

    The full source notes, including the design and AI-method sources, are in SOURCES.md in the repository.