Quant Buffet API

Data API

load_daily_prices, load_price_panel, the disk cache, and price panel conventions.

Market data lives in `backtest.data`. The lab uses adjusted daily closes from Yahoo Finance (auto_adjust=True), cached as CSV under backtest/data_cache/ so repeat runs are fast and reproducible.

load_daily_prices

python

prices = load_daily_prices(
    ["SPY", "TLT", "GLD"],   # symbols: list[str]
    "2010-01-01",            # start: str = "2000-01-01"
    None,                    # end: str | None = None
    use_cache=True,          # keyword-only: read/write backtest/data_cache/
    strict=False,            # keyword-only: raise on first failure
)
# -> pd.DataFrame, DatetimeIndex rows, one float column per loaded ticker

Parameter	Type	Default	Description
`symbols`	`list[str]`	—	Tickers to load, e.g. `["SPY", "TLT"]`.
`start`	`str`	`"2000-01-01"`	Inclusive ISO start date.
`end`	`str \| None`	`None`	Optional end date; `None` means latest available.
`use_cache`	`bool` (keyword-only)	`True`	Read and write the CSV cache.
`strict`	`bool` (keyword-only)	`False`	Raise on the first symbol failure instead of skipping it.

Return value

A `pd.DataFrame` indexed by a sorted, timezone-naive `DatetimeIndex`.
One float column per symbol that loaded successfully — failed symbols are simply absent unless strict=True.
Each column is forward-filled independently, so a symbol's gaps are patched but its pre-IPO history stays NaN.
When some symbols fail, the reasons are attached to `prices.attrs["load_errors"]`.
If no symbol loads, a RuntimeError lists the first few errors.

The disk cache

Aspect	Behaviour
Location	`backtest/data_cache/`
Filename	`{SYMBOL}_{start}_{end or 'latest'}.csv`
Columns	`Date`, `AdjClose`
Reuse condition	File exists and is larger than 50 bytes
Cache key caveat	A different `start` produces a different file — and a fresh download

Note. Because the cache key includes the date range, keeping start="2000-01-01" across experiments reuses one cached file per symbol. Changing start on every run forces new downloads and makes runs noticeably slower.

load_price_panel

python

from backtest.data import load_price_panel

prices, missing = load_price_panel(["SPY", "TLT", "XYZ"], start="2010-01-01")
# prices:  pd.DataFrame of everything that loaded
# missing: ["XYZ"] — symbols with no data

load_price_panel de-duplicates the symbol list, always runs non-strict, and hands back the missing symbols explicitly. Use it in scripts where you want to report gaps rather than crash.

Note. Inside the lab sandbox you normally never call either loader — the runner loads ASSETS for you and passes the panel into make_on_day. Use these APIs for local scripts and batch backtests.

Panel conventions to rely on

Guarantee	Consequence for your code
Index is sorted ascending	`rolling` and `shift` behave as expected
Index is timezone-naive	No `tz_localize` needed for comparisons
Columns follow `ASSETS` order	But may be shorter if a symbol failed
Values are adjusted closes	Splits and dividends are already handled
All-NaN rows are dropped by the runner	The first panel row has at least one real price

python

# Defensive opening lines that belong in every strategy
cols = [c for c in ASSETS if c in prices.columns]
if not cols:
    raise ValueError("No ASSETS loaded")

first_valid = {s: prices[s].first_valid_index() for s in cols}
# Inspect this when a backtest starts much later than you expected