Last Updated: 2026-10-26
A bot is useful only if it performs a defined job within limits you can inspect. “AI”, “free” and a profitable leaderboard do not establish that. First decide whether the job is scheduled buying, a fixed order ladder, testing a coded strategy, or managing many orders. Those require different tools and different supervision.
This guide replaces overlapping rankings and an unsupported manual-versus-AI performance comparison. We do not have the complete funded-trade logs needed to endorse their returns, ratings or survey conclusions. The examples below are synthetic; the provider references describe capabilities, not results we independently tested. Trading and custody can cause losses, including the full amount committed.
Choose the smallest system that does the job
Scroll sideways to see every column. Keyboard: focus the table region and use the arrow keys.
| Job | A starting point to investigate | What you still own |
|---|---|---|
| Make a scheduled purchase | The eligible venue's recurring-buy controls | Purchase size, schedule, fees and continued asset exposure |
| Apply a fixed averaging or grid rule | A native venue tool or documented automation service | Maximum committed capital, inventory, range exits and stop behavior |
| Test a strategy you can express precisely | A research/backtesting framework | Data quality, code, test separation and execution assumptions |
| Quote or coordinate orders across connectors | An execution framework with the required connectors | Inventory limits, stale quotes, reconciliation and operational support |
Adding a paid service before identifying a missing requirement creates another bill and permission boundary. Native tools can remove that extra integration, but cannot make a bad rule safe. A grid can accumulate a falling asset; averaging orders can exhaust reserved capital; leverage adds margin and liquidation risks. See the DCA ladder and grid accounting example for the mechanics.
Three concrete documentation starting points, checked 26 October 2026:
- 3Commas: its v1-to-v2 guide distinguishes DCA setup from TradingView-triggered Signal Bots. An old screenshot or product name is not enough to specify the current workflow. Verify your plan and venue separately.
- Freqtrade: its backtesting reference documents simulation assumptions. Treat it as a research tool, not an independently profitable strategy or a substitute for forward observation.
- Hummingbot: its paper-trade documentation describes simulation and named supported strategy types. Check the exact connector and mode; support for one does not establish support for every venue or strategy.
These are examples of different architectures, not a top-three ranking. No software was installed or connected to a funded account for this review.
“Free” needs a complete cost definition
Distinguish a permanent software licence, a limited plan, a time-limited trial and a venue tool with no separate subscription. Record the end date, automatic renewal, active-bot limit, data entitlement and export rights. A zero subscription does not eliminate exchange fees, spread, funding, hosting or time spent maintaining the system.
Synthetic monthly budget: $10,000 of filled turnover at 0.05% costs $5, not $10. If $10,000 means entries alone and exits add another $10,000, commission is $10. State which definition you use. Add an assumed $20 subscription and $8 hosting: the first case costs $33 before other costs, regardless of whether the strategy earns anything.
On a $500 account, that $33 is 6.6% of starting capital for that month. It is a cost burden, not a required or achievable monthly return. A strategy showing $25 gross profit would still lose $8 after those stated costs. An equally hypothetical two hours of maintenance valued by you at $15/hour adds $30 of opportunity cost, not a cash charge by the venue. Keep those two totals separate.
Do not increase capital, leverage or turnover to make a subscription look proportionally smaller. If the budget does not support the tool, a paper exercise or no subscription is a valid outcome.
Inspect the failure path before the happy path
Use one non-funded test record for each case:
Scroll sideways to see every column. Keyboard: focus the table region and use the arrow keys.
| Failure | Evidence required to continue |
|---|---|
| Submission times out | Existing order state checked before any resend; stable client/order identifiers |
| Only part of an order fills | Filled amount and remaining exposure reconciled; no duplicated full-size replacement |
| Price feed stops updating | Explicit freshness limit; new entries paused instead of reusing an expired signal |
| Process restarts after a loss stop | Halt remains in force until deliberately reviewed; restart does not silently resume orders |
| Connector rejects quantity or price | Precision, minimum size and available balance checked; rejection recorded |
| Exchange or service is unavailable | Open orders and held assets remain visible as unresolved; a retry is not counted as success |
A trading key without withdrawals can still cause losses through unwanted trades. Separate services' credentials and use only the permissions required by the documented integration. Keep secrets out of model prompts, screenshots and exported logs. The workflow comparison separates analysis, execution and custody.
Compare automation with a real manual baseline
Automation can enforce a rule consistently; it can also repeat the wrong rule quickly. A language model summarizing a report is not the same system as software authorized to submit orders. Write down who proposes the action, who validates it, and what is allowed to execute. An assistant's confident sentence is not a validated signal.
To compare methods, choose the same strategy, period, starting exposure, permitted information and cost assumptions. Record manual decisions and missed actions prospectively, not only remembered mistakes. Include monitoring, maintenance and review time on both sides. Do not attribute a result to “AI” when the assets, leverage or market period also changed.
For backtests, separate development from the untouched evaluation interval. Check future-data leakage, missing/delisted assets and sensitivity to nearby parameter values. A beautiful result that disappears after a small fee or parameter change is weak evidence. A forward paper run can reveal operational errors, but does not reproduce real queue position, impact or every outage.
Finish with a dated record: required function, supported product, full cost, permissions, failure cases tested, unresolved inputs and decision. “Cannot establish suitability” is more useful than a forced star score. Keep open losses and failed runs alongside completed wins. This page supplies no minimum starting balance, allocation percentage or expected income.
Disclosure: Educational analysis, not investment advice. AITR may earn commissions on marked affiliate links elsewhere; the documentation links above are references, not evidence of hands-on testing or a recommendation to fund an account.