Backtesting Concepts

This page explains the vocabulary used by FINSABER. If you are new to systematic trading, read this before writing a strategy.

What A Backtest Does

A backtest replays historical market data and asks: "If this strategy had made decisions using only information available at the time, what trades would it have made and what would the portfolio have been worth?"

FINSABER makes the important assumptions explicit:

When the strategy sees information.
When an order is filled.
Which price is used for the fill.
How commission, slippage, liquidity, and LLM cost are charged.
Where outputs are saved for audit and analysis.

Core Terms

Term	Meaning in FINSABER
Bar	One daily price record for a ticker, usually open, high, low, close, adjusted close, and volume.
Signal	A strategy decision, such as "buy AAPL" or "sell MSFT".
Order	The requested trade sent to the framework after a signal.
Fill	The executed trade after timing, price, cash, liquidity, commission, and slippage rules are applied.
Position	Current shares held for a ticker.
Cash	Uninvested portfolio cash after buys, sells, and costs.
Equity curve	Portfolio value through time: cash plus marked-to-market holdings.
Universe	The set of tickers allowed in a run.
Selector	A component that chooses a subset of tickers for a period.
Window	A date range used for one run or one rolling-period evaluation.

Daily Timeline

For date-level data such as daily prices, news, and SEC filings, exact intraday availability is often unknown. The safest default is to generate a signal on day t and trade at the next available open.

sequenceDiagram
    participant D as Data at day t
    participant S as Strategy
    participant E as Execution
    participant P as Portfolio
    D->>S: price/news/filings visible for date t
    S->>E: submit buy or sell signal
    E->>E: wait if execution_timing = next_open
    E->>P: fill at adjusted open on day t+1
    P->>P: update cash, position, costs, equity

Use same_close only when your features are genuinely available before the same day's close. If in doubt, use next_open.

Financial Assumptions

FINSABER separates market assumptions from strategy logic. A strategy should decide what it wants to do; the execution model decides what actually happens.

Assumption	Why it matters
Adjusted prices	Stock splits can make raw prices look like huge losses. Adjusted OHLC keeps simulation prices split-adjusted.
Commission	Every trade costs money; ignoring it overstates high-turnover strategies.
Slippage	The fill price is usually worse than the quoted price, especially for large orders.
Liquidity cap	A strategy should not buy more shares than the market could realistically absorb.
LLM cost	If a strategy pays for model calls, that is an economic cost and can be counted alongside trading costs.

Common Biases

Backtests can look good for the wrong reason

Strong performance is not meaningful unless the data, timing, and universe rules prevent future information from leaking into past decisions.

Bias	Example	Safer practice
Look-ahead bias	Reading tomorrow's filing while trading today.	Use only data available at the decision date.
Same-day text leakage	Using a date-only news article and filling at the same close.	Prefer `next_open` unless timestamps prove availability.
Survivorship bias	Testing only tickers that survived until today.	Use a historically correct universe when possible.
Split error	Treating a 2-for-1 split as a 50% crash.	Use adjusted OHLC for returns and fills.
Liquidity bias	Buying 20% of daily volume at the close price with no impact.	Enable liquidity caps and slippage.
Selection bias	Ranking tickers using the whole test period.	Select tickers using only the training or prior window.

Choosing The Right Engine

Use FINSABERBt when your strategy is naturally written as a Backtrader next() loop or uses Backtrader indicators.

Use FINSABER when your strategy consumes daily dictionaries directly, such as LLM agents that read price, news, filings, and extra modalities.