Appendix D: Quantitative Trading FAQ
This appendix collects the most common confusions and misconceptions from learners, helping you avoid typical cognitive traps.
I. Returns and Risk
Q1: Does a high Sharpe ratio equal low risk?
Not exactly.
Sharpe ratio measures risk-adjusted returns, but has two blind spots:
- Doesn't reflect tail risk: A 2.0 Sharpe strategy may hide 30% maximum drawdown
- Sensitive to leverage: Leverage can artificially boost Sharpe, but risk scales proportionally
Correct approach: Look at Sharpe ratio together with maximum drawdown and Calmar Ratio.
Q2: Backtest shows 50% annual return, how much can I make live?
Rule of thumb: Live returns ~ Backtest returns x 0.3-0.6
Reasons:
- Backtest assumes ideal execution, live has slippage and impact costs
- Backtest has no emotional interference, live has fear and greed
- Backtest environment is stable, live encounters various failures
Correct approach: If backtest returns halved are still acceptable, then consider live trading.
Q3: Why does my strategy backtest great but lose money live?
Three most common reasons:
| Reason | Manifestation | Solution |
|---|---|---|
| Look-Ahead Bias | Used future data | Check signal generation time vs. execution time |
| Overfitting | Training set >> Test set | Walk-Forward validation |
| Cost underestimation | Ignored slippage, impact | Use conservative cost assumptions |
II. Strategy Selection
Q4: Which is better, trend or mean reversion strategy?
There is no "better," only "more suitable."
| Market State | Trend Strategy | Mean Reversion |
|---|---|---|
| Trending market | Big wins | Big losses |
| Range-bound market | Frequent stop-losses | Stable profits |
Correct approach: Identify market regime (Regime Detection), use the right strategy at the right time.
Q5: Can machine learning predict stock price movements?
Almost cannot.
- Financial data has extremely low signal-to-noise ratio, top models only achieve 52-55% accuracy
- 52% accuracy after costs may result in net losses
- Deep learning needs massive data, quant data usually insufficient
Correct positioning: ML is not for "predicting up or down," but for extracting weak but robust signals from noise.
Q6: Are high-frequency strategies easier to make money?
For institutions yes, for retail no.
High-frequency barriers:
- Hardware: Dedicated lines, co-located servers (hundreds of thousands USD/year)
- Software: Sub-millisecond latency (requires C++/FPGA)
- Cost: Being "eaten" by slippage is common in high-frequency
Retail is better suited for: Medium-low frequency strategies (daily/weekly), using cognitive edge rather than speed edge.
III. Risk Control
Q7: What's the right stop-loss level?
No standard answer, but there's a calculation method:
Stop-loss range = k x Volatility
k = 2-3 (too tight gets stopped out easily, too loose loses too much)| Asset Type | Daily Volatility | Suggested Stop-Loss |
|---|---|---|
| Large cap (SPY) | 0.8% | 2-3% |
| Tech stock (TSLA) | 3.5% | 7-10% |
| Cryptocurrency | 5%+ | 10-15% |
Q8: Can diversification eliminate risk?
In normal times yes, in crisis it fails.
| Period | AAPL-MSFT Correlation | Stock-Bond Correlation |
|---|---|---|
| Normal | 0.7 | -0.3 |
| Crisis | 0.9 | 0.6 or -0.5 |
LTCM's lesson: They assumed stable correlation, result was crisis correlations spiked and all assets fell together, diversification failed.
Correct approach: Assume crisis correlations spike to 0.9, stress test with this assumption.
Q9: Can the Kelly formula be used directly?
Cannot use directly, recommend "Half Kelly."
Reasons:
- Win rate and odds are estimates, may be inaccurate
- Full Kelly has too much volatility, psychologically unbearable
- In finance positions are correlated, violating Kelly's independence assumption
Practical recommendation: Kelly/2 as position upper limit, combined with other constraints.
IV. Data and Systems
Q10: Is free data sufficient?
Sufficient for development stage, depends for production.
| Data Type | Free Viability | Recommended Solution |
|---|---|---|
| Daily OHLCV | Sufficient | Yahoo Finance, Alpha Vantage |
| Minute bars | Limited | Broker APIs (requires account) |
| Tick/L2 | Impossible | Must pay (thousands-tens of thousands USD/year) |
Q11: Is Python fast enough?
Sufficient for medium-low frequency, not for high-frequency.
| Strategy Frequency | Latency Requirement | Python Suitability |
|---|---|---|
| Daily/Weekly | Second-level | Fully sufficient |
| Minute-level | Millisecond-level | Marginally usable |
| Sub-second | Microsecond-level | Must use C++/FPGA |
Q12: What to do about API rate limits?
This is a common problem, needs pre-emptive handling.
Solutions:
- Local caching: Store to database after first fetch, reduce API calls
- Rate limiter: Add
time.sleep()in code - Backup data sources: Switch to backup when primary is rate-limited
- Pre-loading: Batch download during non-trading hours
V. Multi-Agent
Q13: Why do we need multi-agent?
Single models cannot adapt to all market regimes.
| Problem | Single Model | Multi-Agent |
|---|---|---|
| Trending -> Range-bound | Strategy fails | Regime Agent switches experts |
| Model overload | Does everything, does nothing well | Expert division of labor |
| Single point of failure | Entire system crashes | Other Agents continue running |
Q14: What if Regime Detection misjudges?
Misjudgment is inevitable, key is controlling consequences.
Design principles:
- Soft switching: Use probability weighting not 0/1 switching
- Confirmation delay: State must persist N days before confirming switch
- Conservative during transition: Reduce positions when uncertain
- Fast crisis response: Rather misjudge as crisis than miss it
Q15: Can Risk Agent be overridden?
Absolutely not. This is a hard constraint of system design.
Risk Agent's veto power:
- Position exceeds limit -> Force reduction
- Drawdown triggers -> Force deleveraging
- Circuit breaker triggers -> Prohibit new positions
Even if Signal Agent has "better reasons," cannot bypass Risk Agent.
VI. Learning Path
Q16: Can I learn quant without programming background?
Can understand concepts, but difficult to practice.
Suggested path:
- First learn Python basics (2-4 weeks)
- Learn pandas/numpy data processing (2 weeks)
- Then learn strategy parts of this course
- Learn by coding and verifying
Q17: How much math background is needed?
High school math + intro statistics is enough to get started.
Core concepts:
- Mean, standard deviation, correlation
- Probability distributions (normal, fat tails)
- Log returns, compound interest calculation
Not needed:
- Advanced calculus (unless doing options pricing)
- Linear algebra (unless going deep into ML)
Q18: How long until I can go live?
Conservative recommendation: 6-12 months.
| Stage | Time | Goal |
|---|---|---|
| Learn basics | 2-3 months | Understand concepts, run backtests |
| Strategy development | 2-3 months | Have a strategy that passes Quality Gate |
| Paper trading | 2-3 months | Verify system stability |
| Small capital live | Ongoing | Use 1-5% of capital to verify |
VII. Common Mindset Issues
Q19: 10 consecutive stop-losses, is strategy failing?
Not necessarily.
Trend strategy typical characteristics:
- Win rate only 30-40%
- Profit/loss ratio 3:1 or higher
- 10 consecutive stop-losses is normal
Judgment criteria: Check if drawdown exceeds 1.5x historical max drawdown. If not, continue executing.
Q20: Can I directly use others' strategies?
Can reference, cannot copy.
Reasons:
- Public strategies are already arbitraged, Alpha has decayed
- Parameters may be overfit to specific period
- You don't understand the strategy's risk characteristics
Correct approach: Understand logic -> Backtest yourself -> Adjust parameters -> Small capital verification.
Key Takeaways From This Appendix
- No Holy Grail: Any strategy will fail at some point
- Risk First: Surviving is more important than making more
- Assumptions Will Be Wrong: Model assumptions fail in extreme situations
- Continuous Learning: Markets change, strategies must evolve with them