by Francis Oyakhire
📂 Full code + backtest receipts: https://github.com/francisx1999/crypto-trading-bot-postmortem
I spent the last six weeks building, deploying, and stress-testing a retail crypto trading bot. Two exchanges, seven distinct strategies, 100 hyperopt epochs, and a 7.3 GB historical dataset later, here's what I have to show for it: every single strategy I tested lost money over the same 24-month window.
This isn't a clickbait hook. The numbers are at the bottom of the post, with the actual freqtrade backtest receipts. What I want to share isn't "I failed" — it's the specific pattern that killed multiple strategy families in identical ways, and the capital math that means even success would barely matter at the scale most retail traders can afford to risk.
If you're considering building one of these, this post is the one I wish I'd read before writing a single line of code.
⚠️ Not financial advice. This is an educational case study in strategies that lost money. Full disclaimer at the bottom.
What I built
Stack: Freqtrade (open-source crypto trading framework), Python, CCXT, deployed on a $5/mo VPS with systemd auto-restart, hardened loopback-only REST API, Telegram alerts, the whole production-grade boring stuff.
Two exchanges over the project lifetime:
- Coinbase Advanced Trade (Apr-May 2026 dry-run) — 0.4% maker / 0.6% taker fees
- Kraken Pro (migration started May 26, 2026) — 0.25% maker / 0.40% taker fees
10-pair USD whitelist: BTC, ETH, SOL, LINK, ENS, AVAX, HBAR, LTC, FIL, ATOM.
This is what a "real" retail crypto bot looks like after you actually do the engineering work. Not a toy.
The kill gate
Before I tested anything, I defined the bar: net ≥ 1.5% per month, compounded, over a 24-month backtest at realistic fees.
Why that bar? Because a $2,000 wallet × 1.5% net per month = $30/month — barely enough to cover a $5 VPS, a $7.99 Claude subscription, and minor tooling. Anything less and the project is structurally underwater on costs. Anything more is meaningful income.
In compounded terms, 1.5%/mo over 24 months = roughly +36% total return. That's the number to clear.
Over 24 months, the market itself (BTC-led basket) was down -13% to -18% depending on weighting. So clearing +36% means the strategy has to net +50pp over what holding would have done. That's a real edge claim, not a marketing one.
Strategy 1: Strict trend-following (v4)
The original strategy. Trend-aligned EMA cross + RSI window + ADX strength filter + MACD confirmation + ATR volatility floor + Fear & Greed sentiment gate. Eight AND-conditions before any entry fires.
Hyperopt-optimized over 26 months of Coinbase backtest. Reported 75% win rate and 1.65% max drawdown.
Deployed live (dry-run) for 44 days on Coinbase. Zero trades fired.
The strategy was working exactly as designed — refusing to trade when conditions didn't meet its (correctly conservative) bar. The problem: my "bar" was so high it almost never triggered. 4 trades in 26 months of historical data. 1.85 trades per year.
When I re-ran v4 against fresh Kraken data:
| Metric | Result |
|---|---|
| Trades in 24 months | 5 |
| Total return | -1.16% |
| Win rate | 60% (3W / 2L) |
| Per-trade expected value | -0.88% |
Even if v4 had traded 30× more often, the per-trade math is negative. Trading more would have lost faster.
Killed by: refusing to trade — and when it did trade, the per-trade edge was negative anyway.
Strategy 2: Loosened trend (v5)
The fix attempt. Same v4 architecture, but loosened: lookback EMA cross trigger (allow entries within a 1-5 bar window of cross, not strict crossover), F&G as position-size scaler instead of hard block, ROI ladder tightened for Kraken's lower fees.
Result on 24 months of Kraken data:
| Metric | Baseline | Hyperopt best (100 epochs) |
|---|---|---|
| Trades | 180 | 428 |
| Total return | -33.08% | -55.74% |
| Max drawdown | 33.9% / 622 days | — |
| Win rate | 51.1% | 41.1% |
180 trades. 0 of 100 hyperopt epochs above the kill gate. Tuning made it actively worse, because the trade-count-rewarding objective just told it to lose money faster.
Killed by: the exit logic. Specifically:
| Exit reason | # | Win% | P&L |
|---|---|---|---|
| ROI take-profit | 88 | 100% | +$740 |
| trailing stop | 4 | 100% | +$37 |
| exit_signal (opposite EMA cross) | 71 | 0% | -$970 |
| stop_loss | 17 | 0% | -$469 |
Read that table carefully. When v5's entries find signal, the ROI exits hit 100% of the time. The strategy is finding real trades. But the exit_signal logic — "close the position when the EMA crosses back the other way" — fires 71 times across the test, every single one of those 71 is a loser, and they collectively bleed more than the wins capture.
This pattern showed up again, identically, in two more strategies below. Hold that thought.
Strategy 3: Grid bot (DCA-on-dips)
Different architecture entirely. Mean-reversion instead of trend-following. Buy on dips below a 20-period moving average, DCA at fixed 1.5% spacing, exit each tranche at +1.5% via ROI ladder. The textbook grid bot mechanic in freqtrade form.
| Metric | Result |
|---|---|
| Trades | 608 |
| Total return | -95.48% |
| Win rate | 93.3% |
| Max drawdown | 96.18% / 642 days |
| Final balance | $2,000 → $90 |
A 93.3% win rate, and you still lost 95% of capital. That's the grid bot trap. 567 small wins (each ROI = +1.5%) totaled +$1,817. But 39 catastrophic stop-losses (-30% each, on a stacked DCA position) bled -$3,700.
The shape: grid bots are short volatility. They quietly accumulate small wins in sideways patches, then get blown up by every trending move. Q2-Q3 2024 SOL crashed -37%, ENS bled -21%, AVAX trended down -15%. Each of those crushed a stack of DCA'd grid tranches.
Killed by: trends. In a market where any pair moves directionally for weeks, grid bots eat losses on every level until stop-loss empties the stack.
Strategy 4: Fear & Greed contrarian
A genuinely different signal class. No EMAs, no RSI, no price action. Pure sentiment: buy when the Fear & Greed Index drops to ≤25 (extreme fear), sell when it rises to ≥75 (greed).
This is the closest thing to a "leading indicator" that retail can access for free. F&G data goes back to 2018 via alternative.me's free API.
| Version | Trades | Total return | Best trade |
|---|---|---|---|
| F&G with -25% stoploss | 28 | -24.39% | SOL +71% |
| F&G with no stoploss | 15 | -17.75% (best result of all tested) | SOL +30% |
Exit breakdown for the stoploss version:
| Exit reason | # | Avg P&L | Total |
|---|---|---|---|
| F&G hit greed (signal exit) | 10 | +32.76% | +$1,012 |
| stop_loss | 13 | -25.59% | -$1,349 |
| force_exit (test ended still-open) | 5 | -9.16% | -$151 |
This was the most interesting finding of the project. When an F&G position survives long enough to ride from extreme fear back to extreme greed, the average profit is +32.76%. Best single trade: SOL/USD +71%. The contrarian thesis works.
What killed it: 13 positions got stopped out at -25% before F&G could rebound. Removing the stoploss helped (-24% → -17%) but introduced the inverse problem — positions ride down 30%+ without protection.
The deeper issue: F&G is fundamentally a BTC sentiment indicator. Applying it as a blanket trigger across 10 alt pairs imports basis risk we never accounted for. A real F&G strategy would probably be BTC-only with position sizing scaled to F&G depth, not naive 10-pair blanket buys.
Killed by: applying a BTC sentiment signal to an alt basket without disciplined risk management.
Strategy 5: Macro regime overlay
The "are we just trading against the macro tape?" hypothesis. Same basic EMA cross entry, but only allow trades when the macro is risk-on (S&P 500 above its 50-day SMA AND Dollar Index below its 50-day SMA).
Over our window the macro was strongly risk-on most of the time (SPX rallied from 4,200 to 6,800; DXY fell from 107 to 98). If the regime filter mattered, this should have been the favorable backdrop.
| Metric | Result |
|---|---|
| Trades | 566 |
| Total return | -57.09% |
| Win rate | 49.1% |
| Max drawdown | 57.6% / 722 days |
Worse than v5. And the exit breakdown was identical in pattern:
| Exit reason | # | Win% | P&L |
|---|---|---|---|
| ROI | 287 | 100% | +$1,356 |
| exit_signal (opposite EMA cross) | 279 | 0% | -$2,498 |
The same systemic killer as v5. Regime filtering changed when the strategy traded but not how it exited. The "exit on opposite indicator cross" pattern bled in identical fashion.
Killed by: the same exit logic that killed v5. Regime context can't fix bad exits.
The pattern across all seven strategies
| # | Strategy | 24mo result | Failure mode |
|---|---|---|---|
| 1 | v4 (strict trend) | -1.16% | Refuses to trade; negative per-trade EV |
| 2 | v5 baseline | -33.08% | Exit-on-opposite-cross logic bleeds |
| 3 | v5 hyperopt best | -55.74% | Tuning can't escape exit logic |
| 4 | Grid bot (DCA) | -95.48% | Trends crush mean reversion |
| 5 | F&G with stoploss | -24.39% | Stop kills winners-in-progress |
| 6 | F&G no stoploss | -17.75% (best loss) | Alt basket can't capture BTC sentiment |
| 7 | Macro regime EMA | -57.09% | Same exit logic as v5 |
Zero of seven cleared the kill gate. Best loss was -17.75% over 24 months — and that was the strategy where I disabled risk management to let positions ride.
The standout diagnostic finding: two of the worst strategies (v5 and macro overlay) failed in identical ways for identical reasons. Different entry logic, different filters, but the same exit pattern — "close when the indicator crosses back the other way" — produces the same shape: 100% win rate on ROI exits, 0% win rate on exit_signal exits, with the latter bleeding more than the former captures.
If you're building a freqtrade strategy with this exit pattern, you have my prediction: it will systematically lose money regardless of entry tuning. I tested it across two architectures, default and hyperopt-optimized, with and without regime filters. Same shape every time.
The capital reality
Even if I'd found a working strategy tonight, the numbers don't pay for the project at retail scale. The math is brutal and worth doing carefully.
A top-decile retail trader (rare and usually survivorship-biased) might hit 30% annualized sustained over years. Let's use that as a generous upper bound.
| Capital | 30% APY net | Per month | Covers a $7.99/mo subscription? |
|---|---|---|---|
| $500 | $150/yr | $12.50/mo | Yes, by $4.51 |
| $1,000 | $300 | $25 | Yes |
| $2,000 | $600 | $50 | Yes |
| $5,000 | $1,500 | $125 | Yes |
| $50,000 | $15,000 | $1,250 | Hits US federal standard deduction (~tax-free income threshold) |
| $100,000 | $30,000 | $2,500 | Married standard deduction territory |
You need around $50,000 of risk capital for a working bot to produce tax-free-income-level returns. Below that, even a winning strategy is essentially a hobby — better than nothing but not life-changing.
The bottleneck isn't strategy edge. It's capital. And trading $2k → $50k requires roughly 25× return; even at a top-decile +5%/month sustained, that's roughly 5 years of perfect execution. Not a realistic path.
What I'd tell my pre-build self
Calculate the capital ceiling before strategy. $2,000 × 30% APY = $50/mo. That's the ceiling, not the expected value. If $50/mo doesn't move your life, the strategy doesn't matter.
The exit logic dominates the entry logic. I spent weeks on entry conditions and hyperopt-tuning them. The thing that killed both v5 and the macro overlay was a one-line "exit when the indicator reverses" decision I'd made in 30 seconds.
Backtest equity curves, not win rates. A 93% win rate that loses 95% of capital is the lesson here. Win rate is a vanity metric. Absolute net-of-fees equity is the only one that matters.
A signal that works on one asset doesn't work on a basket. F&G is fundamentally a BTC sentiment indicator. Treating it as a generic crypto signal across 10 alts imports basis risk that masks whether your signal even works.
Test the alternative gate before declaring trading dead. I almost stopped after v5. The F&G test revealed a +32% average exit when the signal hit. The signal was real; the framework around it was wrong. That's an order of magnitude different from "no signal exists."
Don't confuse activity with progress. I shipped commits, ran backtests, set up infrastructure. None of those things make money. Output is the only metric.
Where this leaves me
Strategy development on this architecture family is closed. Seven strategies, all losing. The diagnostic is rigorous enough that further iteration on EMA-cross/mean-reversion/sentiment-blanket approaches isn't justified at $2k of capital.
What's left, honestly:
- Content like this post — methodology, honest evidence, infrastructure, the failures. The crypto-content space is drowning in "I made $X" claims. This isn't that. It's the actual receipts, with code links.
- Higher-capital research — if you have $50k+ to deploy and want to build something more disciplined, the F&G finding (+32% avg trade) is worth proper isolation. Not on a 10-pair blanket; BTC-only, with position sizing scaled to F&G depth, and exit criteria smarter than "wait for greed." I'm not deploying real capital on that, but it's the one finding that wasn't just "this doesn't work."
- The bigger income lever isn't trading. It's the content, freelance, and consulting work that uses the technical skills the bot built. Bot, grid bot, and stablecoin staking are all variations on the same $10-50/mo lane at retail capital. None of them are the path to meaningful income.
If you're reading this thinking "but I'll find the strategy that works" — maybe you will. But the capital math doesn't care. Run the numbers on your actual deployable capital, multiply by your actual achievable APY, and ask if that monthly number is what you wanted out of the project. If the answer is no, the strategy isn't the bottleneck.
Receipts
All seven strategies, all backtests, all source code is on the project repo: github.com/francisx1999/crypto-trading-bot-postmortem.
Specifically:
-
user_data/strategies/BotTraderStrategy.py— v4 strict trend -
user_data/strategies/BotTraderStrategyV5.py— v5 loosened trend -
user_data/strategies/GridStrategyMinimal.py— grid/DCA -
user_data/strategies/FearGreedStrategy.py— F&G contrarian -
user_data/strategies/MacroRegimeStrategy.py— macro regime overlay - Backtest results are summarized in the tables above; raw logs available on request.
- Data conversion:
scripts/kraken_csv_to_feather.py,scripts/fng_profile.py,scripts/fetch_macro.py
Data sources:
- Kraken official OHLCVT archive (via support.kraken.com)
- alternative.me F&G index (free historical API)
- Yahoo Finance for SPX (^GSPC) and DXY (DX-Y.NYB)
Methodology cross-checks anyone is welcome to run.
Disclaimer
This is for educational and informational purposes only. It is not financial, investment, or trading advice, and nothing in it is a recommendation to buy, sell, or hold any asset.
Cryptocurrency trading carries a substantial risk of loss. Every strategy documented here lost money in backtesting — they are presented as a case study in what does not work, not as a system to deploy. Past performance, including backtest results, does not guarantee future results.
The author is not a licensed financial advisor. Do your own research and consult a qualified professional before making any financial decision. You are solely responsible for any use you make of this code or information, and the author accepts no liability for any losses or damages arising from its use. The software is provided "as is" under the GNU GPL-3.0 License.
USE AT YOUR OWN RISK. THE AUTHOR ASSUMES NO RESPONSIBILITY FOR YOUR TRADING RESULTS.
Built with Freqtrade, Python 3.12, CCXT, deployed on a low-cost VPS. AI-assisted (Claude) for code review, hyperopt loss design, and the postmortem analysis itself.
Copyright © 2026 Francis Oyakhire. Licensed under the GNU GPL-3.0.
