Nessie v2: seeded 'month of transactions' all land on today: create DTOs have no date field and created_at is server time; use a historical simulation for 30 days of history
Symptom: a budgeting app tried to build a month of realistic history by looping over 30 days and POSTing deposits, purchases, transfers and bill payments for each day. Every transaction came back with created_at (and settle_at) of the moment it was created, so the "month" collapsed into a few minutes of one day. Daily aggregation produced a single data point, and the cash-flow forecast step (statsforecast AutoARIMA/AutoETS on a daily series) fell back to a plain average; likely there was simply no daily history to fit.
Cause: the v2 API does not let you backdate. The spec's house rules say timestamps are RFC 3339 UTC, set by the server. The create schemas only accept these fields: - DepositCreateDTO: amount_cents, description, medium - WithdrawalCreateDTO: amount_cents, description, medium - PurchaseCreateDTO: amount_cents, description, medium, merchant_id - TransferCreateDTO: amount_cents, description, medium, payee_id There is no created_at, date or transaction_date field, and the schemas declare additionalProperties: false, so an extra date field will not be honoured (expect a 422).
Options: 1. Use the simulator for history. POST /simulations {"mode":"historical","transactions":2000,"customers":3,"seed":42} writes transactions dated across the past 30 days (failed attempts included). The run object lists its account_ids; filter rows by payer_id/payee_id in that set and sort by created_at, because a historical run writes rows in generation order, not date order. 2. Keep your own "business date" in the description (for example "2026-10-03 groceries") and parse it, accepting that Nessie's timestamps are creation times. 3. Generate the time series locally for forecasting demos and use Nessie only for the live part.
Also note: creating hundreds of transactions in a loop runs into the per-key rate limit (120/min, bursts of 30), so seeding a month by hand is slow as well as undated.
Verify: GET /accounts/<account_id>/transactions?limit=100 and look at the spread of created_at; seeded-by-hand data spans minutes, simulator data spans about 30 days.