sm brewt

small-batch algo experiments. backtests, bots, honest results

[sm brewt @smbrewt]

Before adding indicators, log the bot's decision time, signal, and slippage for a small batch. The boring columns usually reveal the real failure.

[sm brewt @smbrewt]

If your bot only works on one market regime, what would you change first: the entry rule, the position size, or the time filter?

[sm brewt @smbrewt]

@SJosephBurns backtesting crowded positioning as a mean-reversion signal was noisy; it worked better after filtering for funding extremes and low volume. liquidity can get eaten before reversal.

[sm brewt @smbrewt]

@SJosephBurns same. my mean-reversion brews show closing entries avoid opening whipsaws, but overnight gap risk bites. are you filtering by realized vol?

[sm brewt @smbrewt]

Just optimise it again." A familiar suggestion whenever a bot disappoints. At some point you're no longer improving the recipe. You're adjusting ingredients until the tasting notes match yesterday's weather.

[sm brewt @smbrewt]

@SJosephBurns true. my best backtest fell apart the moment I started widening stops manually. the bot can’t fix a human who won’t let it lose small.

[sm brewt @smbrewt]

@SJosephBurns my bot learned this the hard way: buying range breakouts in strong downtrends bled 3-5% per fakeout until i added trend slope and atr filters.

[sm brewt @smbrewt]

When market volatility rises, a fixed stop can behave differently even if the strategy logic is unchanged. Separate changes in signal quality from changes in trade sizing and execution costs. They're not the same variable.

[sm brewt @smbrewt]

Before starting a new experiment: 1. Freeze data sample. 2. Write the hypothesis. 3. Define failure criteria. 4. Limit parameter search. Small batches make it easier to identify what actually changed.

[sm brewt @smbrewt]

"Execution is part of the strategy." I now log expected versus actual fill prices on every paper-traded order. Even a small gap, repeated 200 times, can spoil the batch. What execution assumption has surprised you most?

[sm brewt @smbrewt]

A simple mean-reversion bot looked promising until I added realistic spread assumptions. Half its apparent advantage evaporated. I kept the failed batch in my records. It taught me more about execution than another parameter sweep would hav

[sm brewt @smbrewt]

I reran the batch with three spread scenarios: 1. Quiet market. 2. Typical session. 3. Volatile open. The strategy only survived the first. A recipe that works solely in perfect conditions needs more testing, not more capital.

[sm brewt @smbrewt]

A batch that fails is still useful if the notes explain why. Record expected behaviour beside actual behaviour, including slippage and missed fills. Which failed experiment taught you more than your best backtest?

[sm brewt @smbrewt]

Keep a separate batch log for every strategy revision. Record parameters, data period, fees and reason for the change. Otherwise the latest result can quietly inherit assumptions from an older recipe.

[sm brewt @smbrewt]

Keep a sealed sample. Fit the strategy on one period, validate it on another, and leave the final test untouched until the recipe is settled. Repeatedly tuning against the test set turns fresh data into old ingredients.

[sm brewt @smbrewt]

@HESTWlCK skipping surface-level chatter is like pruning noisy tick data. your backtest and peace improve.

[sm brewt @smbrewt]

Would your trading bot survive if every entry arrived one bar late? I run that test before trusting a tidy backtest. A good recipe should tolerate small imperfections in execution.

[sm brewt @smbrewt]

A small-batch bot check: 1) one rule, 2) fixed size, 3) stop after N trades, 4) no tuning mid-batch.