Quant Study 00 — pandas Basics · Trend Following vs Mean Reversion · The 4 Backtest Biases · Fake Alpha

Starting a new series. Studying quant (quantitative trading). This 00 post covers the basic tools and concepts. Before I run any actual backtest code, I want to organize how pandas handles time-series data and the 4 patterns where backtests often lie.

Table of contents

1. pandas basics — Series and DataFrame

• Series = 1-dimensional (a single row/column)
• DataFrame = 2-dimensional (a table)

A single line of closing price time series is a Series. An OHLCV table (open, high, low, close, volume) is a DataFrame.

loc vs iloc — label-based or position-based

For time series, I usually slice by date using loc. For relative access like “the last N days,” iloc is more convenient.

2. pandas looks at the past — the warm-up period (NaN)

A rolling window operation like a moving average needs N past values to compute. But there’s no past data at the very start of a time series.

• 20-day moving average → on day 1, there aren’t 20 past values, so it’s NaN
• This section is the “warm-up”

How to handle it:

df["MA20"] = df["Close"].rolling(20).mean()
df = df.dropna()  # NaN 행 제거 → 워밍업 잘라내기

dropna() is a method that keeps “only non-NaN values.” If you don’t trim the warm-up, every downstream signal calculation gets contaminated with NaN.

3. Silent data corruption — the trap of adjusted close prices

Even if you pull the same ticker for the same period, the data itself can change the next time you fetch it. This happens because stock splits, dividends, and rights offerings cause historical closing prices to be retroactively re-adjusted (this is the “adjusted close”).

The problem is that this doesn’t throw an error. The numbers just change slightly. So when a backtest that worked fine yesterday suddenly gives different results today, you can spend a long time confused.

How to respond:

• Force a fresh fetch by setting use_cache=False
• For tickers that had a split/dividend event, delete the cache file and re-fetch

A data bug is scarier than a code bug. There’s no error — the returns just look slightly inflated.

4. Strategy classification — trend following vs mean reversion

Even looking at the same price chart, opposite strategies work depending on market regime.

Trend Following

“What’s going up keeps going up”

• Representative indicators: Moving Average (MA), TSMOM (Time-Series Momentum)
• Signal: Golden cross (short-term MA crosses above long-term MA) → buy
• Signal: Dead cross (short-term MA crosses below long-term MA) → sell
• Works well in a trending market

Mean Reversion

“What’s dropped comes back to its place”

• Representative indicators: RSI, Bollinger Bands
• Signal: enter in the opposite direction when far from the mean
• Works well in a sideways market

Key point

You need to switch strategies depending on whether the current market is trending or range-bound. Running mean reversion in a trending market means constantly catching a falling knife, and running trend following in a sideways market means getting chopped up by frequent stop losses.

5. Mismatched warm-up lengths lie to you too

Say Strategy A uses a 20-day MA and Strategy B uses a 200-day MA. If you compare the two over the same period, their warm-ups differ:

• A generates signals starting day 21
• B generates signals starting day 201

A’s first 180 days of trading don’t exist at all for B, so even if you look at the same performance, the starting lines are different. For a fair comparison, you need to trim both strategies based on whichever has the longer warm-up.

Skip aligning the warm-up, and you go straight to a wrong conclusion.

6. The 4 biases of quant — 4 ways backtests lie

(1) Look-ahead Bias

The mistake of making a decision using future data you couldn’t have seen at that point in time.

• Example: Using day t’s closing price to make a buy decision on day t (the close isn’t confirmed until market close, but you’re assuming you knew it during the trading session)
• Example: Backtesting with day t+1’s opening price while “pretending” you decided based on day t’s close

The key to preventing this = a time boundary. Calculate signals only using data up to t-1, and trade on day t.

(2) Overfitting

Results where you’ve fit parameters too well to the past, so they collapse in the future.

• Example: “MA 17-day + RSI 6 + Bollinger 1.8σ” looked fantastic for 2020–2025 → falls apart in 2026
• A combination that fits the past smoothly is likely just fit to the data

Out-of-sample testing can filter this out to some degree.

(3) Unrealistic Trading Costs

Ignoring fees, taxes, and slippage makes returns on paper magically look great. Especially for high-frequency strategies, it’s common for the profit/loss sign to flip after deducting trading costs.

(4) Survivorship Bias

Looking only at surviving stocks means the failed losers are missing. If you calculate historical returns using today’s stock list, companies that have already been delisted after failing aren’t in the dataset. So you end up only covering successful cases, and returns look more inflated than they actually were.

“If a backtest looks too good, be suspicious.” There’s a high chance one of the 4 biases is at play.

Measuring survivorship bias leakage — reviving them yourself

To check how much survivorship bias has crept into your backtest:

Leakage = (number of delisted stocks in the universe) - (number of those that actually made it into the size ranking)

The bigger this gap, the more failed losers were excluded from the size list = the bigger the survivorship bias leakage.

7. Look at the distribution, not a single point — filtering out fake alpha

It’s risky to judge backtest results from a single return curve. Even slightly tweaking parameters or the period for the same strategy can make results swing. So look at a distribution, not a single point.

Not “Did this strategy beat B&H?” (a single point), but “When run across various conditions, is the median alpha positive, and is it reasonable relative to market exposure?” (a distribution)

The trap of “fake alpha”

Even a high B&H beat rate can still be fake alpha. B&H is just a simple buy-and-hold strategy — buy and do nothing. Frequently beating it isn’t necessarily good on its own. Two things you need to check alongside it:

• Median alpha — the median, not the average. It’s common for one or two big wins to pull the average up. If the median is negative, more than half the trades were losses → in live trading, the typical trade loses money.
• Market Exposure — the proportion of time you held a position out of the total period. If a strategy beat B&H while only being exposed to the market 5% of the time, that’s an unfair comparison (mostly holding cash means naturally lower volatility). The shorter the exposure time, the smaller the sample, so there’s a higher chance it won by luck.

Looking only at the B&H beat rate can look great, but there are surprisingly many cases with short exposure time and negative median alpha. “Won often” and “wins consistently” are different things.

Walk-forward Analysis

The standard method for filtering out overfitting. You fix a set of parameters chosen once and roll the time series forward, checking whether it still works in the following period.

[ in-sample : 파라미터 최적화 ] → [ out-of-sample : 그 파라미터로 검증 ]
                ↓ (앞으로 밀기)
       [ in-sample ] → [ out-of-sample ]
                ↓
                ...

• Impressive in-sample but collapses out-of-sample → sign of overfitting
• You need to look at the distribution of out-of-sample results to judge whether it’s real alpha

Distribution shows up again here. A single good out-of-sample result doesn’t make a strategy good — the median/variance of results across multiple sliding windows need to be healthy.

Retrospective — discovering the habit of “benchmarking by feel”

While studying this, I realized I have a tendency to benchmark by feel without even noticing. Something like “using cache feels faster, I think…”

In the example, I actually measured the time and quantified it:

• Network fetch: 0.088 seconds
• Cache read: 0.034 seconds
• Cache is about 3x faster

Once I had the number “3x” in hand, “cache is the right call” finally became a judgment. Before that, it was just an impression. Quant is ultimately a field where you doubt with numbers and decide with numbers, so I think I need to fix this habit first.

Things to study further

1. pandas time series tools

• resample() — converting daily bars → weekly/monthly bars
• shift() — essential for preventing look-ahead bias (signal = df["Close"].shift(1) > df["MA20"].shift(1))
• min_periods option in rolling() — controlling warm-up
• MultiIndex — ticker × date panel data

2. Backtesting frameworks

• backtrader / vectorbt / zipline / bt — building it yourself vs using a framework
• Building it yourself is good for learning, but frameworks are safer for preventing the 4 biases
• Walk-forward optimization, anchored / rolling windows

3. Tools to prevent the 4 biases

• Point-in-time (PIT) data — data exactly as it was known at that point in time (not reflecting later restatements of financial statements)
• Survivorship-bias-free datasets — a universe that includes delisted stocks (commercial data like CRSP)
• Combinatorial Purged CV (Marcos López de Prado) — preventing leakage in time-series K-fold

4. Regime Detection

• Automatically detecting trending / sideways markets → automatic strategy switching
• ADX, Hurst exponent, HMM (Hidden Markov Model)
• “Regime awareness + strategy routing” is more practical than a single strategy

5. Trading cost modeling

• Fees (fixed rate) + taxes (capital gains tax, transaction tax) + slippage (market impact)
• Slippage is proportional to order size relative to trading volume → can’t be ignored once asset scale grows
• If Sharpe drops below 1 after deducting trading costs, hold off on live deployment