| name | tw-stock-quant |
| description | Quantitative investing for Taiwan-listed stocks — backtest framework design, classic factor models (value, momentum, quality, low volatility), strategy development workflow, performance metrics, overfitting and survivorship bias, and post-go-live discipline. Concept-first; pseudo-code examples are language-agnostic (Python / Go / R).
|
| category | finance |
| tags | ["stock","taiwan","tw-stock","quantitative","backtesting","factor-investing","strategy"] |
| keywords | ["backtest","factor model","Sharpe","Sortino","Calmar","survivorship bias","look-ahead bias"] |
| related | ["tw-stock-fundamental","tw-stock-chip","tw-stock-technical","tw-stock-data","tw-stock-trend"] |
Taiwan Stock Quantitative Strategy
Quant isn't really about coding. It's about turning "my investment logic" into rules that are testable, repeatable, and improvable. The code is just the executor.
When to Use This Skill
- Want to systematize and automate manual stock-picking rules
- Want to backtest whether an investment idea worked historically
- Want a factor model that screens monthly
- Want to compare strategies' risk/return characteristics
- Want discipline-driven execution that resists emotional override
1. Core Workflow
1. Hypothesis → "High ROE + low P/E stocks beat the index long-term"
2. Define rules → Pick top 20 of {ROE > 15%, P/E < 15} on day 1 of each month
3. Backtest → Run on 10 years of historical data
4. Evaluate → CAGR, max drawdown, Sharpe
5. Stress test → Sub-periods, parameter shifts, transaction costs
6. Paper trade → 3–6 months on real data without real money
7. Go live → Real money, strict rule execution
8. Periodic review→ Has the strategy stopped working? Has the market changed?
No step is optional. The most common failure is jumping from step 4 to step 7.
2. Backtest Framework
Architecture
┌──────────────────────────────────────┐
│ Backtest Engine │
│ │
│ History → Signals → Sim trade → Stats│
│ │
│ ┌────────┐ ┌────────┐ ┌────────────┐│
│ │ Data │ │ Strategy│ │ Risk mgmt ││
│ └────────┘ └────────┘ └────────────┘│
└──────────────────────────────────────┘
Key components
| Component | Responsibility | Watch out |
|---|
| Data | Provide OHLCV, financials, chip flow | Use adjusted price (dividend-adjusted) |
| Signal | Generate buy/sell/hold per the rules | Don't peek at the future (see below) |
| Simulator | Simulate orders, fills, fees, taxes | TW fees: 0.1425% commission + 0.3% securities transaction tax (0.15% for day-trade) |
| Risk | Stops, take-profit, position caps | Strategy-level risk control |
| Stats | Compute return, drawdown, risk metrics | Compare against benchmark (0050) |
Taiwan-market specifics
- Dividend-adjusted price. Ex-dividend creates price gaps; un-adjusted price drastically overstates returns. Use back-adjusted close.
- Daily price limit. TW has a ±10% daily limit. Backtest must simulate "couldn't buy at limit-up / couldn't sell at limit-down".
- Liquidity. Small caps may average tens of lots/day. Don't simulate orders larger than X% of daily volume (recommend < 10%).
- Survivorship bias. Backtesting only currently listed stocks excludes the ones that delisted = inflated performance. Use a full history including delisted names.
- Earnings release lag. Q1 reports are released by May 15; using Q1 data in April = look-ahead. Use point-in-time data.
3. Common Factor Models
What is a factor?
A factor = a stock characteristic, validated by academia or practice, that explains return differences.
Five classic factors
| Factor | Definition | Logic | TW applicability |
|---|
| Value | Low P/E, low P/B, high yield | Cheap stocks beat expensive ones long-term | High — TW dividend factors work well |
| Momentum | Top performers over the past 3–12 months | Rising keeps rising (trend continuation) | High; pair with stops |
| Quality | High ROE, low debt, stable earnings | Good companies beat bad ones long-term | High |
| Low volatility | Lowest daily-return volatility over 1 year | Low-vol earns better risk-adjusted returns | Mid-high |
| Size | Small market cap | Small caps have historically outperformed | Caution: TW small caps lack liquidity |
Multi-factor composite
Single factors can underperform for stretches. Combining factors:
- Diversifies single-factor risk
- Different factors complement each other in different regimes
- Stabilizes selection
In practice: rank each stock by each factor (percentile), weighted-average, take the top N.
Score = 0.3 × Value + 0.3 × Quality + 0.2 × Momentum + 0.2 × Low-vol
Strategy examples
Dividend strategy
Filter:
1. Listed ≥ 5 years
2. Dividends paid 5 years in a row
3. ROE > 10%
4. Debt ratio < 60%
5. FCF positive
Rank by: 0.5 × yield + 0.3 × ROE + 0.2 × revenue YoY
Hold: top 20
Rebalance: once a year, post ex-dividend
Momentum strategy
Filter:
1. Price above MA240
2. 20-day average volume > 500 lots/day
3. 12-month return > 0
Rank by: 0.6 × 6-month return + 0.4 × 1-month return
Hold: top 15
Rebalance: monthly (1st)
Stop-loss: -10% from entry per name
Value + Quality blend
Filter:
1. P/E < industry average
2. P/B < 2
3. ROE > 15%
4. OCF > Net income
5. EPS growth 3 years in a row
Rank by: PEG ascending
Hold: top 15
Rebalance: after each quarter's earnings
4. Performance Metrics
Return metrics
| Metric | Meaning | Good benchmark |
|---|
| CAGR | Compounded annual growth rate | TW market long-term ~8–10%; strategy should beat market |
| Cumulative return | Total return over the backtest | Watch absolute growth |
| Alpha | Strategy return − benchmark | > 0 to be worth running |
Risk metrics
| Metric | Meaning | Good benchmark |
|---|
| Max drawdown (MDD) | Largest peak-to-trough drop | < 25% good; > 40% dangerous |
| Volatility | Std dev of daily/monthly returns | Lower = steadier |
| Downside risk | Volatility of negative returns only | Reflects pain better than total volatility |
Risk-adjusted return
| Metric | Formula | Good benchmark |
|---|
| Sharpe | (return − risk-free) / volatility | > 1 good, > 1.5 excellent, > 2 elite |
| Sortino | (return − risk-free) / downside risk | A truer quality reading than Sharpe |
| Calmar | CAGR / MDD | > 0.5 acceptable, > 1 excellent |
Practical metrics
| Metric | Meaning |
|---|
| Win rate | Winning trades / total. > 50% is good — but low win rate + high payoff ratio also wins |
| Win/loss ratio | Avg win / avg loss. > 2 is good |
| Trade count | Too few isn't statistically meaningful (target > 100) |
| Turnover | Annual traded value / avg position. Higher = more friction cost |
5. Overfitting — the Quant's Worst Enemy
What it is
Strategy looks perfect in backtest, dies in production.
Why: the strategy learned historical noise, not pattern.
Warning signs
| Sign | Note |
|---|
| Backtest > 30% CAGR with almost no drawdown | Too good, probably wrong |
| Hyper-precise parameters (MA17, RSI(13.5)) | Why 17 and not 15 or 20? |
| Performance collapses with small parameter shifts | Brittle = overfit |
| Only works in a specific time window | Lucky alignment |
| Many conditions (> 5 filters + 3 indicators) | Each condition adds an overfitting dimension |
Defenses
- Out-of-sample testing. Develop on the first 70% of data, validate on the last 30%. If validation drops sharply → overfit.
- Walk-forward. Rolling backtest: train on past N years, validate next year, roll forward.
- Parameter robustness. Vary key parameters within a sane range (MA20 → MA15..MA25); performance shouldn't shift dramatically.
- Simple beats complex. Fewer conditions, fewer parameters, more intuitive logic = less overfitting risk.
- Economic logic. A strategy needs a reasonable economic story. "Low P/E beats long-term" has a story (value premium); "buy stock #38 on day 17" doesn't.
6. Strategy Development Workflow
Step 1: Hypothesis
Write one sentence: "I think ____ stocks will ____ because ____."
For example:
- "High-ROE + low-P/E stocks beat the index because the market underprices quality."
- "Stocks with 3 consecutive monthly revenue highs rise short-term because of the momentum effect."
If you can't fill in "because", the hypothesis isn't worth backtesting.
Step 2: Pseudo-code rules
On the first trading day of each month:
1. Get all listed and OTC stocks
2. Exclude: listed < 1 year; 20-day avg volume < 200 lots; financials
3. Filter: ROE > 15%; P/E < industry average; revenue YoY > 0 for 3 months
4. Rank: PEG ascending
5. Take top 15
6. Equal weight (1/15 each)
7. Compare to last month's holdings; rotate
8. Per-name stop: -15% from entry
Step 3: Backtest
- Period ≥ 5 years (covering at least one bull/bear cycle)
- Include transaction costs: 0.1425% commission + 0.3% securities transaction tax = ~0.44% round-trip
- Benchmark: TAIEX or 0050 ETF
- Log per-year return, MDD, holdings
Step 4: Evaluate
Compare strategy vs benchmark on:
- CAGR
- Max drawdown
- Sharpe
- Yearly win/loss
A strategy that beats the market with 40%+ MDD will probably break the operator's nerves — not a good strategy in practice.
Step 5: Stress test
- Restrict the period (e.g., only the financial-crisis window, only COVID)
- Vary parameters (MA20 → MA15/MA25; 10/20 holdings)
- Inflate costs (assume 0.1% slippage)
- Tighter liquidity (exclude < 500 lots/day)
If any stress test takes the strategy from green to red → it isn't robust.
Step 6: Paper trade
Run on real-time data, follow rules, don't actually order. 3–6 months. Goals:
- Validate viability in real market conditions
- Find issues backtests miss (limit-up, illiquidity)
- Build execution discipline
Step 7: Go live
- Start small (20–30% of total capital)
- Strict rule following — no human override
- Log every trade, including any deviation from the rule
- Review monthly / quarterly
Step 8: Periodic review
Strategies can fail when market structure changes. Ask:
- Is the last 6 months' performance materially off from backtest?
- Does the factor logic still hold? (Value can underperform during tech bubbles, for instance.)
- Is this a regime change or just a low patch?
A strategy that trails the market for 1 year + with broken economic logic → consider pause or revision.
7. Practical Constraints in Taiwan
| Constraint | Impact | Counter |
|---|
| Small caps illiquid | Backtest looks great, you can't actually buy | Exclude < 500 lots/day average |
| Daily price limit | Model says buy; market is at limit-up | Simulate limit-up / limit-down |
| Ex-dividend price moves | Un-adjusted price = distorted return | Use adjusted close |
| Earnings release lag | Need point-in-time data to avoid look-ahead | Strictly use data on its release date |
| No share borrow | Shorting is hard in TW | Focus on long-only |
| Fees + taxes | ~0.44% per round-trip; high turnover bleeds | Lower turnover |
8. Common Traps
| Trap | Note | Counter |
|---|
| Survivorship bias | Only currently-listed stocks | Use full history with delisted |
| Look-ahead bias | Using data not yet released at the time | Strict point-in-time data |
| Data mining bias | Test 100 strategies, pick the best | Hypothesize first, then test |
| Ignoring fees | Backtest without commission/tax | Always include |
| Ignoring slippage | Filling at close is idealized | Add 0.05–0.1% slippage assumption |
| Over-optimization | Tune to perfection | Robustness check |
| Sample too small | Backtest with 30 trades | At least 100 for statistical significance |
9. Pre-Live Checklist
For each strategy before live deployment:
Related Skills
