Maximum drawdown

How it works

The mechanism is a running maximum over a value series. At each observation, divide the value by the highest value seen so far and subtract one; that is the drawdown at that point, zero at a new high and negative below it. ffn's documentation puts it as current / hwm - 1, the high-water mark being the running peak. The maximum drawdown is the most negative of those numbers over the whole series. Portfolio Visualizer's documentation defines it as the maximum observed loss from a peak to a trough before a new peak is attained, and abbreviates it MDD.

Three properties follow from the arithmetic alone, and they are what make two figures incomparable:

Sampling can only make it shallower. Month-end values are a subset of the daily closes. Every peak-and-trough pair visible in the monthly series is also in the daily one, and the daily one has pairs the monthly series does not. So on the same history a monthly maximum drawdown is at most as deep as a daily one, a year-end figure at most as deep as a monthly one, and a close-to-close figure at most as deep as one that saw the intraday lows.

Length can only make it deeper. Extending the series adds pairs and removes none, so a longer history's maximum drawdown is at least as deep as a shorter history's inside it.

It is one event. The number describes the worst path segment in the sample, not a distribution, and a single observation cannot say how likely a repeat is. The companion figures — how long the fall took, how long recovery took, and the whole underwater period from peak to new peak — are separate numbers that Portfolio Visualizer reports beside it.

Why it matters here

The allocation backtesters in retirement planning and the libraries next to them compute this differently, and read against their own documentation and source code on 26 September 2026 the differences fall into four places.

Sampling frequency. Portfolio Visualizer calculates maximum drawdown from monthly portfolio balances. Portfolio Charts measures drawdowns with year-end data and says so, adding that losses in the middle of a year may have been deeper than shown. testfolio works on daily values — its average drawdown is taken across all days below an all-time high. Backtesting.py and Backtrader record equity on every bar, and both value open positions at that bar's close, so on daily bars an intraday low that recovered by the close is not in the figure.

Nominal or real. Portfolio Charts adjusts all returns for inflation, expressed in the home country's currency. A real drawdown in a high-inflation decade is deeper than the nominal one other tools print for the same years.

Equity or price. ffn's drawdown functions take a price series and empyrical-reloaded's max_drawdown a returns series; hand either an asset and you get the asset's drawdown. The backtesters measure account equity, cash included, so a strategy that is half in cash draws down less than the asset it trades. Where money moves in and out, Backtrader measures total net asset value by default and can switch to a fund-style value in which cash added or withdrawn changes a share count instead; only the second keeps a withdrawal from reading as a loss — the same separation a time-weighted return makes. QuantStats accepts prices or returns and decides which it was given from the values: a series whose maximum is below 1, or whose minimum is zero or less, is read as returns and compounded, so a price series that never reached 1 is not measured as a price.

Sign and units. empyrical-reloaded, ffn and QuantStats return a negative fraction, -0.35. Backtesting.py reports Max. Drawdown [%] as a negative percentage. Backtrader's max.drawdown is a positive percentage, and QuantConnect's LEAN engine stores the absolute value of a fraction. A spreadsheet that collects these side by side and sorts them has to normalise first.

The data underneath decides the rest. A drawdown over a universe of companies that still exist omits every fall that ended at zero, which is survivorship bias at its most literal. And a drawdown read off a stitched walk-forward record depends on whether the equity curve runs continuously across the segments or restarts at each one.