# Test report: risk, not return

Created on 2026-09-16 09:22 UTC with trader 0.1.0 · generated by `scripts/risikobericht.py`

Deutsche Fassung: [risk-portfolio.de.md](risk-portfolio.de.md)

## Question

Every earlier measurement set the bot's return against buy & hold - and it
lost every time. The question was wrong: with crypto, buy & hold also means
riding every crash all the way down. This report measures both, return and
drawdown, on the same capital and with the same fees.

## Verdict

> What holds in every period tested: the bot caps the loss from peak at 21 % at most, while simply holding goes through at least 59 %. What does NOT hold: that it earns more. Of 4 start years tested, it is ahead of holding in 1 - namely when starting 2022-01, shortly before the crash. Anyone starting later sees a trend follower lose clearly: starting 2023, it was +77 % against +527 %. The bot trades return for calm, and that trade only pays off if a crash comes. On the 714 bars of the exchange that is actually traded on: +2.8 % against +31.4 % when holding, with 3.1 % instead of 29.9 % drawdown.

## Long series: 9 markets, 2022-01-01 to 2026-09-13

Computed over the period all markets have in common.

| Variant | Return | Max. drawdown | Sharpe | Trades | Fees | Return per drawdown point |
|---|---|---|---|---|---|---|
| **Bot** (strategy, stops, kill switch) | +66.4 % | -20.9 % | 0.74 | 261 | 19.90 EUR | 3.17 |
| Holding, same exposure (52.3%) | +23.4 % | -41.0 % | 0.30 | 0 | 0.32 EUR | 0.57 |
| Holding, fully invested | +44.7 % | -77.1 % | 0.46 | 0 | 0.62 EUR | 0.58 |
| Holding, BTC only | +64.0 % | -67.2 % | 0.46 | 0 | 0.62 EUR | 0.95 |

## The same bot, started later

The table above starts shortly before the 2022 crash - the period that suits a
trend follower with stops best. Starting later paints a different picture.
Hence the same measurement from four start dates:

| Start | Bot | Drawdown | Holding, full | Drawdown | BTC only | Drawdown |
|---|---|---|---|---|---|---|
| 2022-01 | +66.4 % | -20.9 % | +44.7 % | -77.1 % | +64.0 % | -67.2 % |
| 2023-01 | +77.4 % | -16.0 % | +526.9 % | -59.3 % | +364.5 % | -53.4 % |
| 2024-01 | +42.2 % | -16.2 % | +455.7 % | -63.1 % | +81.4 % | -53.4 % |
| 2025-01 | +11.6 % | -15.5 % | +165.1 % | -63.5 % | -18.2 % | -53.4 % |

## Short series: 9 markets, 2026-05-17 to 2026-09-13

These are the markets and the exchange the bot actually trades on. The
exchange's interface offers no more history than this.

| Variant | Return | Max. drawdown | Sharpe | Trades | Fees | Return per drawdown point |
|---|---|---|---|---|---|---|
| **Bot** (strategy, stops, kill switch) | +2.8 % | -3.1 % | 0.81 | 14 | 0.96 EUR | 0.90 |
| Holding, same exposure (52.3%) | +16.4 % | -16.3 % | 1.64 | 0 | 0.32 EUR | 1.00 |
| Holding, fully invested | +31.4 % | -29.9 % | 1.68 | 0 | 0.62 EUR | 1.05 |
| Holding, BTC only | -1.4 % | -24.3 % | 0.07 | 0 | 0.62 EUR | -0.06 |

## How it was tested

| | |
|---|---|
| Strategy | `sma_cross {'fast': 50, 'slow': 200, 'mode': 'ema', 'trend_filter': 0, 'long_only': True}` |
| Timeframe | 4h |
| Starting capital | 200 EUR |
| Fee / slippage | 26 bps / 5 bps, also when holding |
| Long series | ETH/USDT, BTC/USDT, XRP/USDT, SOL/USDT, ZEC/USDT, UNI/USDT, ADA/USDT, NEAR/USDT, AAVE/USDT |
| Data checksum, long | `0f2df3225956` (10,299 bars, 2022-01-01 to 2026-09-13) |
| Short series | ETH/EUR, BTC/EUR, XRP/EUR, SOL/EUR, ZEC/EUR, UNI/EUR, ADA/EUR, NEAR/EUR, AAVE/EUR |
| Data checksum, short | `a40875cd7118` (714 bars, 2026-05-17 to 2026-09-13) |

"Holding" buys once, equally weighted, and never touches anything again.
"Same exposure" invests the same share of the account as the bot and keeps
the rest in cash - so it is visible what is due to the strategy and what to
the smaller amount of capital deployed.

## What this report does NOT show

* **The parameters were chosen once, not re-fitted per window.** A
  walk-forward of the same strategy kept only about a third of its training
  performance in the test. The figures above are the optimistic reading.
* **The start date matters.** See the table with four start years. A single
  number from a single series says little.
* **The long series is a different exchange with different markets.** It says
  something about the method, not about this very account.
* **The short series is too short for a verdict.** It shows what the account
  goes through, not what the strategy is worth.
* **Not a forecast.** Past prices allow no reliable conclusion about future
  ones.

*Technical evaluation on historical data, not investment advice.*