We did not make money in July: a drawdown of −3.46% across the accounts the bot manages. Below is why it happened, what was rewritten during the month and what we learned. No smoothing over: the channel and the website publish real trades, so a report on a bad month is just as mandatory as one on a good month.
Two thirds of this drawdown is a single position that is still open. The month's closed trades came out close to zero: on price movement the result was positive, but it was eaten by fees and one stuck trade that was still open at the time of publication.
Why the month ended at zero
July was the most intense month of development in the bot's history. This was not cosmetics: entries, exits, result accounting and the evaluation method itself were all rewritten. When you rebuild the foundation, it is more honest to treat that period's return as a cost rather than a failure.
Four new entry filters
Dead market, a slow-drift zone, a pause during a Bitcoin rally, a hollow reversal. Each one is not an idea out of thin air but the result of analyzing hundreds of real trades. They all share one purpose: do not enter where the move will not start.
The filter evaluation mechanism was rewritten
Previously the first filter to trigger stopped the check, and the statistics credited it with others' work. Now filters vote independently, and the value of each is measured by its unique contribution: if two filters cut an entry, neither gets the credit — turn either one off and the trade still won't pass.
It turned out that 42% of the cut signals are caught by more than one filter. In other words, half of the “effectiveness” was double counting.
Targets rebuilt at the “knee”
We found pairs where the take-profit sat at +1% while the price kept going about +3% after the exit. Too cheap a target with a wide stop is a risk-to-reward ratio the pair does not deserve.
Five accounting and execution bugs fixed
A position closed by the exchange at take-profit could hold its slot for up to 15 minutes. Fills on a “wick” slipped past the price polling. Some closes left stale records that caused a trade's result to be lost. This is not strategy, it is hygiene — but without it every measurement lies.
Testing the filters: history and live numbers
Each filter was run over the whole database of real trades since April, not over a pretty simulation. Applied to this history, the full set filters out about two thirds of entries, and among the remaining ones the share of stuck trades falls roughly by half — from 15% to 8%.
But history is still history. So since launch every skipped signal is recorded and played out: what would have happened if the trade had been taken. By the day of publication, 40 such skips had matured:
- 25 would have reached the target
- 14 would have hung until the trade expired
- 1 would have hit the stop-loss
- in total the skips saved 8 percentage points
The key point in these numbers is this: the share of stuck trades among the cut ones is 35% versus the usual 15%. More than double. So the filters do not hit at random but exactly at the class of trades they were built for — the ones that open, go nowhere and hang until they are force-closed.
And right away, what we cannot say yet: that there are already fewer stuck trades in real trading. Less than a week has passed and there are only 37 live trades — any conclusion at this volume would be self-deception. The number keeps being tracked and will go into the August report.
The reverse is visible too: three new filters already pay for themselves, one is in the red for now — it cut trades that would have reached the target. The decision on it will be made in August, based on data rather than on fondness for our own idea.
What the month revealed is worth more than profit
The backtest lied
The simulation showed a 76% win rate; reality gave 61%. The reason is simple: in the simulation a trade has three outcomes, but in live trading it has seven — every fifth trade was closed by rules that did not exist in the test. Now every rule is checked on real trades, not on history.
Losses come from time, not from the stop
The price reaches deep stop levels in only about 3% of cases. Money is lost differently: a position hangs until it expires and is closed at market. This turned our whole approach to exits around.
A rising Bitcoin is more dangerous than a falling one
We expected exactly the opposite. Entering a falling market turned out fine, but a Bitcoin rally hits both sides at once — longs and shorts. In a rally altcoins trend, and reversals in them get trampled.
42% of trades turn against us in the first three hours
Their win rate is 35% versus 80% for the rest. None of the methods we tested can tell in advance which trade will end up in this half. It is an honest open question, not a solved problem.
Macroeconomics does not break the strategy
One hypothesis was that stuck trades are caused by Fed meetings and inflation and jobs data releases. It did not hold up on hundreds of trades: such events widen the range of the move but do not set its direction.
What did not make it into the system
More was rejected this month than was deployed. That is not a weakness of the approach — it is the approach.
- Extending the trade's lifetime “let it recover” — by our measurements a coin flip, and it works worst exactly when it seems the drop has stopped.
- Moving the position to breakeven — the sign of the result flipped from month to month.
- Early cutting of fading trades — turned off after testing: 11 of the 20 cut trades would have reached the target. The rule was cutting maturing entries, not junk.
- Machine-learning trade selection — on new data the prediction was no better than a coin toss.
What's next
The foundation for measurement is in place. Every filter is monitored daily, every skipped entry is played out and compared with what would have happened without it. A rule that stops being useful is now visible within days, not quarters — one such rule has already been turned off, and another is in question.
The month did not bring money. But it brought what money cannot be counted without: a system that shows what really works and what only seemed to.
Try CryptoLogix
An AI agent for auto-trading on Binance Futures. Start with signals and enable auto-trading when you are ready.
Launch the agent →Signals channelFAQ
Why publish a report if the month was a loss?
Because the channel and the website show the account's real trades, not hand-picked examples. Reporting only good months is advertising, not reporting. A losing month with an analysis of the causes is more useful to the reader than a nice number without an explanation.
What does the −3.46% drawdown mean?
It is the change in the accounts' value for July relative to the deposited capital, including fees and open positions. Two thirds of this figure comes from one trade that is still open, so the result is not final.
What are entry filters and why are there so many?
They are rules, each of which can cancel a trade before it opens. A momentum-reversal signal appears often, but most such reversals are false. Each filter cuts a specific class of bad entries: a dead market, an exhausted bounce, a move without volume, and so on.
Why are fewer trades a good thing?
Because in this strategy losses come not from stop-losses but from trades that opened and went nowhere. Each such position takes up a slot in the portfolio and is closed by time. Cutting them at entry is cheaper than closing them later.