Andes Retail — units sold, next 28 days

Six stores, ten SKUs, one line per store/SKU/day. Planning wants a unit forecast for every row of data/holdout.csv, which is the 28 days that follow the last day in data/sales.csv.

Time: 90 minutes. Write everything in solution.py.

Files

Columns

Known defects, stated as rules

1. Re-synced lines. Some row_id values appear twice, as identical copies.
2. Refunds. A refund is booked as a negative units_sold against the same line. Refunds are not demand.
3. One store was closed. For a stretch of consecutive days one store has no rows at all — not rows with zero units. Nothing marks this; you have to find it.
4. unit_price and promo_flag need coercing before anything numeric works.

Tasks

T1 — parse_dates(df) -> pd.DataFrame (10 pts)

Return the frame with sale_date as datetime64. Nothing may fail to parse, and the earliest and latest dates must be the real ones — getting day and month the wrong way round still produces valid dates, which is why this is graded on the range and not on the dtype alone.

T2 — parse_prices(df) -> pd.DataFrame (10 pts)

Return the frame with unit_price numeric. One value per input row, nothing dropped.

T3 — sales_summary(df) -> dict (15 pts)

Given a date-parsed frame, return exactly these four keys:

T4 — clean(df) -> pd.DataFrame (15 pts)

Return a frame with the duplicated lines removed, the refund rows removed, unit_price numeric, and promo_flag filled with 0 where it was blank.

T5 — split_last_days(df, days) -> (train, valid) (15 pts)

Split a date-parsed frame in two by time: valid is the last days calendar days inclusive, train is everything before. Every row goes to exactly one side, and the last training date must be strictly earlier than the first validation date.

T6 — fit_predict(train, holdout) -> np.ndarray (35 pts)

Both frames arrive raw, exactly as the CSVs read. Return predicted units_sold for every row of holdout, in file order.

Graded on RMSE: <= 14.0 beats predicting a single number, <= 9.00 beats a store×SKU historical average, <= 6.50 means price and promotion are in the model too.

Allowed: pandas, numpy, scikit-learn.

Visible cases — Andes Retail — daily demand

What CodeSignal shows next to Run tests: small inputs with the output they should produce. Run them with make drill-test. They use toy data, never the day's CSVs, so passing them is not a grade — make drill-grade is.

T1 — parse_dates (Two date formats in one column)

ISO and DD/MM/YYYY read as the same dates

Input:

 sale_date store_id     sku
2026-03-04       S1 SKU-001
04/03/2026       S4 SKU-001
2026-01-31       S1 SKU-001
31/01/2026       S4 SKU-001

Expected:

sale_date datetime64 == [2026-03-04, 2026-03-04, 2026-01-31, 2026-01-31]

T2 — parse_prices (Prices that arrived as money)

the symbol and the thousands separator both go

Input:

 sale_date store_id     sku  unit_price
2026-03-04       S1 SKU-001 S/ 1,234.50
2026-03-04       S1 SKU-002     S/ 9.90
2026-03-04       S1 SKU-003    S/ 12.00

Expected:

unit_price == [1234.5, 9.9, 12.0]

T3 — sales_summary (Four facts about the book)

refunds, gross units, and the store with missing days

Input:

 sale_date store_id     sku  units_sold
2026-01-01       S1 SKU-001           3
2026-01-02       S1 SKU-001          -1
2026-01-03       S1 SKU-001           4
2026-01-04       S1 SKU-001           2
2026-01-05       S1 SKU-001           5
2026-01-01       S2 SKU-001           2
2026-01-02       S2 SKU-001           6
2026-01-05       S2 SKU-001           7

Expected:

{'refund_rows': 1, 'gross_units': 29, 'closed_store': 'S2', 'closed_days': 2}
(S2 has no rows on the 3rd and 4th; -1 is not demand)

closed days are calendar days, not missing rows

Input:

 sale_date store_id     sku  units_sold
2026-02-10       S3 SKU-001           1
2026-02-11       S3 SKU-001           2
2026-02-12       S3 SKU-001           3
2026-02-13       S3 SKU-001           4
2026-02-10       S5 SKU-001           5
2026-02-13       S5 SKU-001          -2
2026-02-10       S3 SKU-002           1
2026-02-11       S3 SKU-002           2
2026-02-12       S3 SKU-002           3
2026-02-13       S3 SKU-002           4
2026-02-10       S5 SKU-002           5
2026-02-13       S5 SKU-002          -2

Expected:

{'refund_rows': 2, 'gross_units': 30, 'closed_store': 'S5', 'closed_days': 2}
(S5 is absent on the 11th and 12th: two days, four missing rows)

T4 — clean (Clean it)

duplicate line out, refund out, promo blank is 0, price numeric

Input:

row_id  sale_date store_id     sku  unit_price  promo_flag  units_sold
    R1 2026-01-01       S1 SKU-001     S/ 5.90         1.0           3
    R2 2026-01-01       S1 SKU-002 S/ 1,010.00         NaN           2
    R3 2026-01-01       S2 SKU-001     S/ 5.90         0.0          -1
    R2 2026-01-01       S1 SKU-002 S/ 1,010.00         NaN           2
    R4 2026-01-01       S2 SKU-002     S/ 9.50         NaN           4
    R5 2026-01-01       S2 SKU-003    S/ 12.00         1.0           0

Expected:

4 rows (R1, R2, R4, R5); no negative units; promo_flag [1, 0, 0, 1]; unit_price float, max 1010.0

T5 — split_last_days (Hold out the last 21 days)

the last 3 calendar days go to valid

Input:

 sale_date store_id     sku  units_sold
2026-01-01       S1 SKU-001           0
2026-01-02       S1 SKU-001           1
2026-01-03       S1 SKU-001           2
2026-01-04       S1 SKU-001           3
2026-01-05       S1 SKU-001           4
2026-01-06       S1 SKU-001           5
2026-01-07       S1 SKU-001           6
2026-01-08       S1 SKU-001           7
2026-01-09       S1 SKU-001           8
2026-01-10       S1 SKU-001           9
days = 3

Expected:

train: 2026-01-01..07 (7 rows); valid: 2026-01-08..10 (3 rows)

a day with three rows is still one day

Input:

 sale_date store_id     sku  units_sold
2026-01-01       S1 SKU-001           0
2026-01-02       S1 SKU-001           1
2026-01-03       S1 SKU-001           2
2026-01-04       S1 SKU-001           3
2026-01-05       S1 SKU-001           4
2026-01-06       S1 SKU-001           5
2026-01-07       S1 SKU-001           6
2026-01-08       S1 SKU-001           7
2026-01-09       S1 SKU-001           8
2026-01-10       S1 SKU-001           9
2026-01-10       S1 SKU-001          10
2026-01-10       S1 SKU-001          11
days = 3

Expected:

valid: 5 rows over 2026-01-08..10 (the 10th appears three times); train: 7 rows

T6 — fit_predict (Forecast the next 28 days)

one non-negative forecast per holdout row

Input:

160-row toy history (2 stores x 2 SKUs x 40 days, raw export strings, one refund); 20-row raw holdout for the next 5 days

Expected:

array of 20 finite floats, none negative