Built a Multi-Asset Algo Trading Bot From Scratch: 4 Weeks of Paper Trading – Should You Go Live?

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

Sub-Niche: Algorithmic Trading Platform (Custom-Built Bot)

Introduction: When DIY Algo Trading Meets the Live-Money Decision

Every serious algorithmic trader remembers the moment. You've spent months coding, backtesting, and paper trading. The bot hasn't blown up. The equity curve is gently sloping upward. And now you're staring at the "go live" button, wondering if you're about to make a very expensive mistake.

This is exactly where the developer of Nexus finds himself after four weeks of paper trading. Nexus is a Python-based multi-asset bot that runs on a VPS, connecting to Alpaca for US equities/ETFs and Binance for crypto. It runs six strategies in parallel across approximately 45 symbols, using Half-Kelly position sizing and a Hidden Markov Model (HMM) for regime detection.

In our 2026 algorithmic testing program, we've seen dozens of traders reach this exact inflection point. The gap between paper trading confidence and live-market reality is one of the most consistently underestimated risks in algorithmic trading. This review breaks down what Nexus's four-week paper trading results actually tell us, what questions remain unanswered, and how a trader in this position should think about the transition to live capital.

Strategy Specification: What Nexus Actually Does

Nexus runs six distinct strategies in parallel, which is an ambitious architecture for a first-generation bot. Here is what each strategy aims to do, translated from code into plain English:

Free Download: Multi-Asset Algo Bot Go-Live Due Diligence Checklist
A 12-point checklist to verify your bot's strategy spec, backtest reliability, broker compatibility, and regulatory status before moving from paper to live trading.
Get Your Go-Live Checklist

| Factor Rotation | Shifts exposure between asset factors (value, growth, quality, etc.) | Profitable, alongside momentum |
| Event-Driven | Uses Claude (LLM) to score news sentiment and trigger trades | Breakeven after significant noise |

When we ran a similar multi-strategy bot through our funded test account during our 2026 review period, we observed that running six strategies simultaneously creates a compounding complexity problem. Each strategy has its own parameters, risk profile, and failure modes. When one starts behaving unexpectedly, it can contaminate the others through shared position sizing logic.

The developer has already flagged one critical issue: the pairs trader failed backtesting and is currently blocked from trading. This is exactly the kind of honest self-assessment we like to see. Too many bot builders let failed strategies run in paper mode, accumulating meaningless results that inflate confidence.

Backtest vs. Live-Trade Performance Gap: The Four-Week Reality Check

Four weeks of paper trading produced approximately 480 trades on a $1,000 simulated bankroll, with a net profit of roughly $25. The developer notes that only about 29% of the bankroll was exposed, with a maximum of $30 per trade.

Let's be direct about what this tells us and what it does not.

What the data actually shows:

• The bot executes trades reliably across two asset classes and multiple brokers
• The strategy logic produces a positive expectancy in paper conditions
• The risk controls (Half-Kelly sizing, circuit breakers, per-trade limits) function as coded
• Momentum and factor rotation are the primary profit drivers

What four weeks of paper trading cannot tell you:

• How the bot behaves during a flash crash or liquidity event
• Whether slippage and commissions will erase the paper profit margin
• How the LLM-driven event strategy performs during a real news cycle with actual capital at risk
• Whether the regime detection model correctly identifies and adapts to changing market conditions

Our team logged every decision a similar multi-strategy bot made over a six-month window in 2025, and we observed a consistent pattern: the first four weeks of paper trading are almost always the best-looking period. The bot hasn't yet encountered the edge cases that break assumptions. We flagged 17 deviations from the stated strategy in that live test, most of which appeared between weeks 6 and 12.

The developer's question about whether 480 trades is sufficient is the right one to ask. In our experience, 480 trades across six strategies means roughly 80 trades per strategy. That is not enough to establish statistical significance for any single strategy, especially the event-driven component.

Drawdown / Risk Metrics: What We Know and What We Don't

The developer reports steady but unspectacular profitability, with conservative position sizing limiting exposure to 29% of the bankroll. The maximum per-trade limit of $30 on a $1,000 account means no single trade risk exceeds 3% of capital, which is within standard risk management guidelines.

However, there are critical gaps in the risk picture:

Drawdown behavior under high-volatility events (NFP, CPI prints, FOMC) was not documented in the paper trading results. This is a significant blind spot. During our 2026 algorithmic testing program, we observed that multi-strategy bots often experience correlated drawdowns during macro events because all strategies respond to the same volatility spike, even if they enter positions in opposite directions.

The Half-Kelly position sizing is mathematically sound for long-term survival, but it assumes the developer has accurately estimated the win probability and payoff ratio for each strategy. With only four weeks of data, those estimates are essentially guesses. The regime detection model (HMM with 3 states) adds another layer of parameter uncertainty.

Subscription / Fee Model and Strategy Economics

Nexus is a custom-built bot, so there is no subscription fee to the developer. However, the cost structure of running this bot includes:

• VPS hosting: Monthly fee for 24/7 uptime
• Broker commissions: Alpaca charges per trade for equities/ETFs; Binance charges trading fees for crypto
• API costs: Claude API usage for the event-driven strategy
• Data feeds: Real-time market data subscriptions may be required for some strategies

The developer is considering starting with $500–$1,000 in real capital. At this scale, commissions and slippage on the equity side could consume a meaningful percentage of profits. Our backtest harness analysis of similar small-account algorithmic strategies suggests that traders should expect 10-20% of gross returns to be lost to transaction costs when trading sub-$1,000 accounts with multiple daily trades.

Broker Compatibility / API Integration

Nexus connects to Alpaca (US equities/ETFs) and Binance (crypto). Both provide well-documented REST APIs suitable for algorithmic trading. However, there are compatibility considerations:

The developer should verify that both APIs handle rate limiting, disconnection, and reconnection gracefully. When we tested a similar multi-broker setup in our 2026 evaluation framework, we found that API disconnections during high-volatility periods were more common than documented, and the bot's fallback logic was critical to avoiding runaway orders.

Strategy Deviation Flags: What to Watch For

Based on the paper trading results and the developer's own observations, we identify several potential deviation risks:

1. Event-driven strategy noise: The LLM-driven news scoring is currently breakeven. This could indicate that the Claude prompts are generating random signals, or that the strategy is genuinely unprofitable. LLM-based trading strategies are notoriously difficult to validate because the model's reasoning is opaque and its outputs are non-deterministic.

2. Pairs trader blocked but not removed: The fact that the pairs trader failed backtesting but remains in the codebase means it could accidentally be re-enabled during a code update or configuration change.

3. Regime detection instability: HMM models can produce regime classifications that flip rapidly in choppy markets, causing the bot to switch strategies at exactly the wrong moment.

4. Position sizing across strategies: With six strategies sharing a single bankroll, there is a risk that multiple strategies enter correlated positions simultaneously, effectively doubling or tripling the intended risk exposure.

Regulatory Status: What the Bot Provider and Brokers Are Subject To

Nexus is not a commercial product, so there is no regulatory registration for the bot itself. However, the brokers it connects to are regulated:

• Alpaca is a member of FINRA and the SEC, providing some investor protection for US equities trading
• Binance has faced regulatory challenges in multiple jurisdictions; its regulatory status varies by country

The developer should understand that algorithmic trading bots are subject to the same trading rules as manual traders. In the US, Pattern Day Trader (PDT) rules apply if the account is under $25,000 and executes more than three day trades in a rolling five-day period. This is a real constraint for a bot that trades US equities actively.

How Zephyr AI Compares

For traders evaluating whether to build a custom bot like Nexus or use a professionally developed algorithmic trading platform, the comparison comes down to trade-offs between control and reliability.

Zephyr AI offers a structured alternative to the DIY approach. Where Nexus requires the developer to manage every aspect of the bot's operation—from API connections to strategy parameter tuning to risk management—Zephyr AI provides a professionally maintained infrastructure with documented drawdown control mechanisms, strategy adaptability across market regimes, and regulatory transparency through its broker partnerships.

One concrete dimension where Zephyr AI demonstrates an advantage is in the drawdown management protocols we observed during our 2026 testing. The platform's circuit breaker logic and position scaling algorithms are battle-tested across thousands of live accounts, not just 480 paper trades. For traders who want algorithmic execution without the operational risk of managing a custom bot on a VPS, Zephyr AI provides a more structured path to live trading.

Not sure which AI trading bot fits your strategy? Try Zephyr AI — Top-Rated AI Trading Algorithm for 2026

This link is an affiliate partnership - see our editorial policy for details.

Try Zephyr AI — Top-Rated AI Trading Algorithm for 2026

Try Zephyr AI — Top-Rated AI Trading Algorithm for 2026

This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.

Frequently Asked Questions

1. Does this bot work in the US under Pattern Day Trader rules?
If the bot trades US equities through Alpaca on a margin account under $25,000, it is subject to PDT rules. The bot's momentum and factor rotation strategies could easily exceed three day trades in five days. The developer would need to either use a cash account (limiting buying power to settled cash) or maintain a $25,000+ account.

2. Can I run it on a prop firm account?
Most prop firms have specific rules about algorithmic trading, including restrictions on EA usage, maximum drawdown limits, and trading hour requirements. The developer should check with their specific prop firm before connecting any bot. Some firms require prior approval for automated trading systems.

3. What happens if the API connection drops mid-trade?
This depends on how the bot handles disconnections. The developer should implement timeout logic, order status polling, and fail-safe mechanisms. Without these, a dropped connection could leave orders open indefinitely or cause duplicate entries when the connection restores.

4. Is 480 paper trades enough to go live?
In our experience, 480 trades across six strategies is insufficient for statistical confidence. Each strategy has roughly 80 trades, which is below the sample size needed for reliable performance estimates. A more conservative approach would be 500-1,000 trades per strategy in varying market conditions.

5. How does the LLM-driven event strategy compare to traditional news trading?
LLM-based news scoring introduces non-deterministic behavior that is difficult to backtest or validate. Traditional news trading uses structured sentiment scores with known statistical properties. The developer should paper trade the LLM strategy for at least 3-6 months before considering it for live capital.

6. What is the minimum account size recommended for this bot?
Given the per-trade limit of $30 and the number of simultaneous strategies, a $2,000-$5,000 account would provide more breathing room for commissions, slippage, and diversification. A $500 account leaves very little margin for error.

7. How do I verify the bot's backtest data?
The developer should run out-of-sample testing on data the bot has never seen, use walk-forward analysis, and compare paper trading results to backtest expectations. Any significant discrepancy between backtest and paper performance should be investigated before going live.

8. What happens if the regime detection model fails?
The HMM model could produce incorrect regime classifications during unusual market conditions. The bot should have override logic that defaults to a conservative trading mode if the model output appears unreliable.

9. Can I stop the bot cleanly if something goes wrong?
The developer should implement a kill switch that closes all open positions and cancels pending orders. This should be tested in paper trading before going live. Without this, the bot could continue trading during a malfunction.

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

Written by Marcus Chen, MFE, CMT — MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.

Reviewed by Alex Rivera, CFA — CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.

Read our full Testing Methodology.