Generation STEM
Learning LibraryQuant LibraryTeens

Backtesting: study a rule on past data

Would a rule have worked? A backtest runs a strategy over past data to study how it behaved. It's a science experiment, not a crystal ball. (Educational only — never financial advice.)

The big idea

Backtesting replays a trading rule over historical data to measure how it would have performed — a study tool, not a prediction.

See it in code

1The basics

The simplest backtest is one line of study: pretend you bought on day one and held to the end, then measure the change. That's the baseline every fancier rule gets compared against:

python
# Simplest backtest: buy on day 1, hold to the end - study only.
closes = [100, 98, 102, 105, 103, 108]

start = closes[0]
end = closes[-1]
ret = (end - start) / start * 100
print(f"Buy and hold: {ret:.1f}%")
Run it — how buy-and-hold would have done on this data:
Buy and hold: 8.0%

Up 8% over the whole window — but that ignores when the price moved. A smarter rule reacts to the day-to-day changes, so first we need to see them.

2A step further

Step through the days and compare each close to the one before — the day-over-day move that any rule watches. No money yet, just reading the sequence:

python
closes = [100, 98, 102, 105, 103, 108]

for i in range(1, len(closes)):
    change = closes[i] - closes[i - 1]
    direction = "up" if change > 0 else "down"
    print(f"Day {i + 1}: {closes[i]} ({direction})")
Run it — each day labelled up or down versus yesterday:
Day 2: 98 (down)
Day 3: 102 (up)
Day 4: 105 (up)
Day 5: 103 (down)
Day 6: 108 (up)

Day 3 is the first up day. A rule can act on exactly that moment — so next we let the backtest buy there and hold.

3In our world

Now put money on the rule: buy once on that first up-day, then hold. We track cash and shares and measure the final value — purely to observe, not to recommend:

python
# A backtest runs a rule over PAST data to study it - not advice.
closes = [100, 98, 102, 105, 103, 108]

cash = 1000.0
shares = 0
for i in range(1, len(closes)):
    if closes[i] > closes[i - 1] and shares == 0:
        shares = cash / closes[i]
        cash = 0.0
        print(f"Day {i + 1}: BUY at {closes[i]}")

value = cash + shares * closes[-1]
print(f"Final value: {value:.2f} (started at 1000)")
Run it — the rule buys once, and we measure the result:
Day 3: BUY at 102
Final value: 1058.82 (started at 1000)

It bought on Day 3 — the first up-day we spotted above — and held to the end. The final value tells us how this rule did on this frozen data — nothing more. A real backtest tests many rules over long histories, always aware that the past does not promise the future.

The same idea, everywhere

Replaying a decision rule over recorded data to evaluate it is a general method: testing a cache policy on request logs, a routing rule on traffic history, a game-AI on past matches. It's controlled experimentation — change one rule, rerun, compare.

Try it yourself

Add a sell rule (sell when the price falls) and see how the final value changes. Then compare two different rules on the same data — which behaved better, and why?

The common mistake

Believing a good backtest guarantees future profit. It doesn't — markets change, and a rule tuned to fit past data ('overfitting') often fails on new data. Backtesting is for understanding a strategy's behavior, not for making real money decisions.

What it unlocks

Backtesting combines market data, conditionals, and trading signals, and is measured with risk basics.

Related concepts