Can AI Really Do Quantitative Research?

This isn't hype, it's fact.

In this experiment, I used 5 prompts to have AI complete the entire workflow from "stock screening" to "backtesting" to "optimization". The final strategy achieved:

48.17% Annualized Return
1.69 Sharpe Ratio
-30.38% Max Drawdown

In this article, I'll demonstrate 5 types of quantitative research tasks AI can perform, and share the optimization process for the final strategy.

Note: The following are historical backtest results and do not represent future performance.

Skill 1: Revenue New High Screening

First task: Find stocks with revenue hitting new highs.

Find stocks where the 3-month average revenue hits a 12-month high, with sufficient liquidity (20-day average volume > 500 lots)

      
        
        Show Code
        
      
      from finlab import data
from finlab.backtest import sim
 
# Get revenue and volume data
rev = data.get("monthly_revenue:當月營收")
vol = data.get("price:成交股數") / 1000  # Convert to lots
 
# Calculate conditions
rev_ma3 = rev.average(3)
rev_high = (rev_ma3 == rev_ma3.rolling(12).max())
vol_filter = vol.average(20) > 500
 
# Select stocks meeting conditions
position = rev_high & vol_filter
 
# Backtest
report = sim(position.loc['2015':], resample='M', position_limit=0.1, upload=False)
report.to_html('skill1_revenue_high.html')
    

Pure revenue new high achieves about 19.77% annualized. A decent start, but not good enough.

Skill 2: Institutional Flow Analysis

Second task: Find stocks with consecutive trust fund buying.

Find stocks with trust fund buying for 5 consecutive days, select top 10 by buying volume

      
        
        Show Code
        
      
      from finlab import data
from finlab.backtest import sim
 
# Get trust fund trading data
trust_net = data.get("institutional_investors_trading_summary:投信買賣超股數")
vol = data.get("price:成交股數") / 1000
 
# Trust fund buying for 5 consecutive days
trust_buy_5d = (trust_net > 0).sustain(5)
 
# Liquidity filter
vol_filter = vol.average(20) > 300
 
# Combine conditions
cond = trust_buy_5d & vol_filter
position = trust_net[cond].is_largest(10)
 
# Backtest
report = sim(position.loc['2015':], resample='M', upload=False)
report.to_html('skill2_trust_buy.html')
    

Trust fund strategy needs to be combined with other conditions to work effectively.

Skill 3: Technical Indicator Strategy

Third task: Build a momentum strategy with RSI and MACD.

Build a momentum strategy using RSI and MACD:

RSI > 50 indicates bullish

MACD golden cross

Price above 60-day moving average

      
        
        Show Code
        
      
      from finlab import data
from finlab.backtest import sim
 
close = data.get("price:收盤價")
vol = data.get("price:成交股數") / 1000
 
# Technical indicators
rsi = data.indicator("RSI", timeperiod=14)
macd, macd_signal, macd_hist = data.indicator("MACD", fastperiod=12, slowperiod=26, signalperiod=9)
sma60 = close.average(60)
 
# Condition combinations
cond_rsi = rsi > 50
cond_macd = (macd > macd_signal) & (macd.shift() < macd_signal.shift())
cond_trend = close > sma60
vol_filter = vol.average(20) > 300
 
# Entry signals
cond = cond_rsi & cond_macd & cond_trend & vol_filter
position = close[cond].is_largest(20)
 
# Backtest
report = sim(position.loc['2015':], resample='W', upload=False)
report.to_html('skill3_technical.html')
    

Pure technical indicator strategy shows unstable performance, needs fundamental support.

Skill 4: Value Investing Screening

Fourth task: Find undervalued quality companies.

Find undervalued stocks: P/E < 15, Dividend Yield > 5%, ROE > 10%

      
        
        Show Code
        
      
      from finlab import data
from finlab.backtest import sim
 
# Get valuation data
pe = data.get("price_earning_ratio:本益比")
div_yield = data.get("price_earning_ratio:殖利率(%)")
roe = data.get("fundamental_features:ROE稅後")
vol = data.get("price:成交股數") / 1000
 
# Value screening conditions
cond_pe = (pe > 0) & (pe < 15)
cond_yield = div_yield > 5
cond_roe = roe > 10
cond_vol = vol.average(20) > 300
 
# Combine conditions and select stocks
cond = cond_pe & cond_yield & cond_roe & cond_vol
position = div_yield[cond].is_largest(15)
 
# Backtest
report = sim(position.loc['2015':], resample='M', upload=False)
report.to_html('skill4_value.html')
    

Value investing strategy achieves about 13.19% annualized. Stable but not exciting.

Skill 5: Consecutive Condition sustain()

This is a unique FinLab feature: sustain() detects whether conditions hold consecutively.

      
        Show Code
        
      # Revenue YoY growth > 10% for 3 consecutive months
rev_growth = data.get("monthly_revenue:去年同月增減(%)")
consecutive_growth = (rev_growth > 10).sustain(3)
 
# This is stricter than single conditions, filtering out one-time growth

This feature plays a crucial role in the ultimate strategy.

The Problem: No Single Strategy Is Good Enough

After seeing four skills, what did you notice?

Strategy	Annualized Return	Problem
Revenue New High	~20%	Doesn't filter false breakouts
Trust Fund Buying	Unstable	Needs fundamental support
Technical Indicators	Unstable	Too much noise
Value Investing	~13%	Lacks momentum

No single strategy reaches the 30%+ target.

This is where AI's true value begins...

AI Quantitative Research: 5 Prompts to Build a 48% Annual Return System