"진짜 되는" 투자 전략 찾기: 금융전략과 통계적 검정

2016년 7월 23일, 퀀트 스터디 그룹 발표 자료

Transcript

1. "૓૞ غח" ై੗ ੹ۚ ଺ӝ Әਲ਼੹ۚҗ ా҅੸ Ѩ੿ July 23, 2016 नߧળ © Quant Study, Beomjun Shin, July 23, 2016
2. ௫౟ ୡࠁ੄ ੌ઱ੌ 1. (ಂఠ࠙ࢳ) ࣻহח ߎਸ ࢜ݴ ߸ࣻܳ ݃ҳ ߸ഋ೧ࠄ׮ 2. Ӓۧѱ ݆ࣻ਷ ੹ۚ(?)ਸ ٜ݅ѱ ػ׮ 3. ߮஖݃௼ب ޅ੉ӝח ੹ٜۚী ઝ੺ೠ׮ -> द੢੉ ബਯ੸ੌѢঠ..(?) 4. যו ࣽр 10֙ ־੸ ࣻ੊ܫ 160% ੿ب ҡଳই ࠁ੉ח ੹ۚ ߊѼೠ׮ 5. (оࢸѨ੿&प೷ࢸ҅) ৬৬! Ӕؘ ੉Ѣ ޺Ҋ ై੗೧ب غա...??? © Quant Study, Beomjun Shin, July 23, 2016
3. (ಎ؊ݭణ) ߸ࣻܳ যڌѱ ࢎਊೞ૑? [߸ࣻ]ܳ [઱ӝ]۽ [߸ഋ]೧ ࠁҊ [ҙ੼]ীࢲ [৘ஏӝр]݅ఀ ࠁਬ • ߸ࣻ: Value, Momentum, Quality • ઱ӝ: 1ѐਘ, 1௪ఠ(=3ѐਘ), 1֙, 3~5֙ • ߸ഋ: пઙ ੉زಣӐ, ݽݭథ, Valua9on Model • ҙ੼: Loser-Follow(contrarian), Winner-Follow • ৘ஏӝр: 1ѐਘ, 1௪ఠ(=3ѐਘ), 1֙, 3~5֙ © Quant Study, Beomjun Shin, July 23, 2016
4. ਋৬! ҡଳই ࠁ੉ח ੹ۚ੉׮! © Quant Study, Beomjun Shin, July 23, 2016
5. Mul$ple Tes$ng Problem © Quant Study, Beomjun Shin, July 23, 2016
6. प೷ ௏٘ SP500_YEARLY_VOL = 0.15 LENGTH = 252*10 SIZE = 100 DATEINDEX = pd.date_range(start="2011-01-01", periods=LENGTH) normal_walks = pd.DataFrame() for n in range(SIZE): normal_walks[n] = pd.Series( (np.random.normal( loc=0, scale=SP500_YEARLY_VOL / np.sqrt(252), size=(LENGTH, 1) ) + 1).cumprod(), index=DATEINDEX ) -> ࣻ੊ܫ 160%ח 100ߣ ೧ࠁݶ 1ߣ਷ աৢ ࣻ ߆ী হ׮!1 1 ࣻ੊ܫ੉ Normalਸ ٮܲ׮Ҋ о੿, 10֙ زউ ై੗ೠ׮Ҋ о੿, S&P 500 ࣻળ੄ Vola0lityܳ ࢎਊೣ © Quant Study, Beomjun Shin, July 23, 2016
7. ೧Ѿ ߑߨ ? © Quant Study, Beomjun Shin, July 23, 2016
8. Sta$s$cs & Good Experiment Design! © Quant Study, Beomjun Shin, July 23, 2016
9. ݾର • оࢸѨ੿җ प೷ࢸ҅ • In-Sample ࠙ࢳ • p-value ৬ Mul/ple Tes/ng Problem ೖೞӝ • In/Out-Sample ࠙ࢳ • ಂఠ ࠙ࢳ • ੸੺ೠ ߮஖݃௼ ࢸ੿ • N࠙ਤ ࠙ࢳ प೯ © Quant Study, Beomjun Shin, July 23, 2016
10. ా҅೟ ӝࠄ ਊয • ా҅೟(sta&s&cs)? • ಴ࠄ(sample)۽ࠗఠ ݽ૘ױ(popula&on)ਸ ਬ୶ೞח Ѫ • ా҅۝(sta&s&c)? • ݽ૘ױ੄ ౠ૚ਸ ୶੿ೞӝ ਤ೧ ࢠ೒۽ࠗఠ ҅࢑ೞח ӏ஗ • ಴ࠄ࠙ನ(sampling distribu&on)? • ా҅۝(sta&s&c)੄ ഛܫ࠙ನ • ಴ࠄ࠙ನо ઺ਃೞ׮. © Quant Study, Beomjun Shin, July 23, 2016
11. ా҅೟ োण • ୡҗ ࣻ੊ܫ(Excess Return)ী ؀ೠ ୶੿ਸ ೧ࠁ੗ • о੿1: ୡҗ ࣻ੊ܫ਷ ੿ӏ࠙ನܳ ٮܲ׮ • о੿2: ࢠ೒੄ ࣻ(୨ োب੄ ࣻ) n੉ ੘׮ • ഛܫ߸ࣻ • ా҅۝(sta,s,c) ח ੗ਬب ੋ t-࠙ನ(಴ࠄ࠙ನ;sampling distribu,on)ਸ ٮܴ ੉ ঌ۰ઉ੓׮ • ҙण੸ਵ۽ Әਲ਼ীࢶ t-sta,s,c੉ 2ܳ ֈযঠ ࢎਊೞחѦ۽ ߉ই٘۰૓׮ © Quant Study, Beomjun Shin, July 23, 2016
12. ੿ӏ࠙ನܳ ٮܰחо? © Quant Study, Beomjun Shin, July 23, 2016
13. p-value = type 1 error def. sta)s)c > cri)cal value or significance level > p-value © Quant Study, Beomjun Shin, July 23, 2016
14. power = 1 - type 2 error def. sta)s)c > cri)cal value © Quant Study, Beomjun Shin, July 23, 2016
15. Type 1 Error, Type 2 Error © Quant Study, Beomjun Shin, July 23, 2016
16. Low p-value or High p-value in Stock Return? © Quant Study, Beomjun Shin, July 23, 2016
17. T ా҅۝җ ࢥ೐࠺ਯ 2 2 r਷ yearly return੉ۄ о੿, daily return਷ ܳ ғೞৈ োਯചػ ࢥ೐࠺ਯ۽ ࢎਊ. ਷ ݽ؛੄ Return, ח ߮஖݃௼੄ Return © Quant Study, Beomjun Shin, July 23, 2016
18. ୡҗࣻ੊਷ ઓ੤ೞחо? • ӈޖоࢸ: ୡҗࣻ੊੉ হ׮ ( ) = 'ژ੉ ੹ۚ'੉ۄ ೞ੗ • ؀݀оࢸ: ୡҗࣻ੊੉ ੓׮ ( ) = 'ঌ౵ ੹ۚ'੉ۄ ೞ੗ © Quant Study, Beomjun Shin, July 23, 2016
19. ҡଳই ࠁ੉ח(?) ੹ۚ੄ оࢸѨ੿ from scipy.stats import t nyears = 10 yearly_sharpe = 0.75 tstat = yearly_sharpe * np.sqrt(nyears) pvalue = 1 - t.cdf(tstat, df=9) print(tstat, pvalue) # => 2.37170824513 0.0208959003886 • p-value: 2% • Ѣ૙ ೧ࢳ: ঌ౵੹ۚ੉ۄҊ ೠ ੉ ੹ۚ੉ ૓૞ ঌ౵੹ۚੌ ഛܫ੉ 98% ੉׮ • ଵػ ೧ࢳ1: ژ੉੹ۚੋؘ ঌ౵੹ۚ੉ۄҊ ೡ ഛܫ੉ 2% ੉׮ • ଵػ ೧ࢳ2: ژ੉੹ۚੋؘ ژ੉੹ۚ੉ۄҊ ೡ ഛܫ੉ 98% ੉׮ © Quant Study, Beomjun Shin, July 23, 2016
20. Base Rate=0.1੄ ੄޷ 10ѐח ૓૞ ঌ౵੹ۚ, 90ѐח ૓૞ ژ੉੹ۚ © Quant Study, Beomjun Shin, July 23, 2016
21. Sta$s$cal Power=0.8 ੄ ੄޷ 10ѐ ૓૞ ঌ౵੹ۚ ઺ ୶ஏ ঌ౵੹ۚ਷ 8ѐ © Quant Study, Beomjun Shin, July 23, 2016
22. P-value=0.02 ੄ ੄޷ 90ѐ ژ੉੹ۚ ઺ ୶ஏ ঌ౵੹ۚ਷ ѐ © Quant Study, Beomjun Shin, July 23, 2016
23. Python Code N = 100 base_rate = 0.10 alpha = 0.02 beta = 0.2 contingency_table = pd.DataFrame( [ [ N*base_rate*(1-beta), # pred ঌ౵, real ঌ౵ N*base_rate*beta # pred ঌ౵, real ژ੉ ], [ N*(1-base_rate)*alpha, # pred ژ੉, real ঌ౵ N*(1-base_rate)*(1-alpha) # pred ژ੉, real ژ੉ ] ], index=pd.Series(["ঌ౵੹ۚ", "ژ੉੹ۚ"], name="True"), columns=pd.Series(["ঌ౵੹ۚ", "ژ੉੹ۚ"], name="Pred")) fdr = contingency_table.iloc[1, 0] / contingency_table.iloc[:, 0].sum() print("FDR:", fdr) contingency_table © Quant Study, Beomjun Shin, July 23, 2016
24. Con$ngency Table © Quant Study, Beomjun Shin, July 23, 2016
25. P-value © Quant Study, Beomjun Shin, July 23, 2016
26. Power © Quant Study, Beomjun Shin, July 23, 2016
27. False Discovery Rate © Quant Study, Beomjun Shin, July 23, 2016
28. P-valueী ؀ೠ ೧ࢳ ׮द ੿ܻ • [ঌ౵੹ۚ੉ۄҊ ೠ ੹ۚ੉ ژ੉੹ۚੌ ഛܫ]੉ [2%] ੉׮ [X] • [ঌ౵੹ۚ੉ۄҊ ೠ ੹ۚ੉ ژ੉੹ۚੌ ഛܫ]਷ False Discovery Rate(FDR)੉޲۽ ੉׮ [O] • [ژ੉੹ۚਸ ঌ౵੹ۚ੉ۄҊ ೡ ഛܫ]਷ [2%]੉׮ [O] • FDRਸ ஶ౟܀ೞח Ѫ੉ ઺ਃ೤פ׮. © Quant Study, Beomjun Shin, July 23, 2016
29. Mul$ple Tes$ng Problem ࢎ۹1: ੐੄੄ ౟ۨ੉٬ ੹ۚী 100ѐ੄ ےؒ ߸ചܳ ઱Ҋ ਗې ੹ۚҗ P&L ਸ 2-sample t-test ೠ׮Ҋ ೞ੗. 100ѐ੄ ےؒ߸ചо ੹ࠗ ޖ੄޷ೠ ࢚ട (base rate = 0)੐ীب p-valueܳ 5% ࣻળਵ۽ ౸ױೞݶ 5ѐח ਬ੄ೞѱ աৢ Ѫ੉׮. Ӓۧ׮ݶ Ӓ 5ѐ੄ ੹ۚਵ۽ ై੗ೞݶ غחо? NO! ࢎ۹2: ୭Ҋ੄ ై੗੗ غח ߑߨ, 10000ݺ੄ Ҋёীѱ Ӓ ઺ ੺߈਷ ղੌ ௏ झೖо য়ܲ׮Ҋ ੉ݫੌਸ ࠁղҊ աݠ૑ ੺߈਷ ߈؀۽ ৘ஏ Ѿҗܳ ࠁմ׮. ੉Ѧ 10ߣ ߈ࠂೞݶ 10ݺ੄ ҊёীѲ ղо 10ߣ ݽف ݏ୸ ࣅ੉ ػ׮. դ ୭ Ҋ੄ ై੗੗ੋо? NO! © Quant Study, Beomjun Shin, July 23, 2016
30. ೧Ѿೞח 2о૑ ҙ੼ • Family-Wise Error Rate(FWER) • ੺؀੸ੋ ҙ੼ীࢲ ױ 1ѐ੄ type 1 errorо ߊࢤೞח Ѫب ೲਊೞ૑ ঋח׮ • 100ѐ੄ పझ౟ա 10000ѐ੄ పझ౟ա ױ 1ѐ੄ type 1 errorب ߊࢤೞݶ উػ׮ • False Discovery Rate(FDR) • ࠺ਯ ҙ੼ীࢲ type 1 errorо ߊࢤೞחѦ ೲਊೞ૑ ঋח׮. • 100ѐ੄ పझ౟ীࢲ 1ѐܳ ೲਊೞ૑ ঋח׮ݶ 10000ѐ੄ పझ౟ীࢶ 100ѐܳ ೲ ਊೞ૑ ঋח׮ © Quant Study, Beomjun Shin, July 23, 2016
31. FWER: Bonferroni’s Correc2on • 1ѐ੄ పझ౟: p-value 5% • 10ѐ੄ పझ౟: p-value 5% -> 5/10 % • 100ѐ੄ పझ౟: p-value 5% -> 5/100 % ࢤп. • ై੗੹ۚ਷ ইಫ۽ ೐۽ં౟о ইפӝ ٸޙী ઑӘ ਬೞѱ ࠊب غ૑ ঋਸө? • ݅ড Nߣ੄ పझ౟о completely uncorrelated੉ݶ ਤ୊ۢ ೧ب ؼ Ѫ੉׮, ೞ૑ ݅ perfectly correlated غ੓׮ݶ ࢎप࢚ పझ౟ח Nߣ੉ ইפۄ 1ߣೠ ࣅ੉ ػ׮ © Quant Study, Beomjun Shin, July 23, 2016
32. FWER: Holm-Bonferroni Method • ծ਷ p-value ࠗఠ ֫਷ p-value ࣽਵ۽ ੿۳ • kח పझ౟੄ ੋؙझ, M਷ ੹୓ పझ౟ പࣻ, \alphaח significance level • sta6s6c੄ ࠙ನܳ ࢎਊೞח ѐ֛ © Quant Study, Beomjun Shin, July 23, 2016
33. FDR: Benjamini and Hochberg, Yeku9eli(BHY) • ֫਷ p-value ࠗఠ ծ਷ p-value ࣽਵ۽ ੿۳ • о੢ וटೠ ߑߨ • harvey৯, о੢ ੸੺ೞ׮Ҋ ঌ۰૓ ߑߨ © Quant Study, Beomjun Shin, July 23, 2016
34. ਃড • • Mul%ple Tes%ng Problem & Solu%on • Interpret p-value © Quant Study, Beomjun Shin, July 23, 2016
35. Walk-Forward Tes/ng © Quant Study, Beomjun Shin, July 23, 2016
36. Terms • (Sta&s&cs) In-sample(IS) Data / Out-of-sample(OOS) Data • (Machine Learning) Training Data / Valida&on Data / Test Data / Cross-Valida&on © Quant Study, Beomjun Shin, July 23, 2016
37. Avoid Overlap Bias © Quant Study, Beomjun Shin, July 23, 2016
38. അप੄ प೷ ࢸ҅: ೟ण੉ হח ݽ؛ 1. ୭Ӕ N֙ ؘ੉ఠܳ Out-of-sample(OOS) ؘ੉ఠ, աݠ૑ܳ In-sample(IS) ؘ੉ ఠ۽ ܻ࠙೧ك׮ 2. IS ؘ੉ఠܳ о૑Ҋ प೷ਸ ߈ࠂೠ׮ • प೷੄ ߈ࠂ਷ ઱۽ ݽ؛੄ ߸ച, ౵ۄ޷ఠ੄ ߸ച۽ ҳࢿػ׮ • ಌನݢझܳ ҙ଴ೠ പࣻܳ ߈٘द ӝ۾ೠ׮ 3. о੢ જও؍ TOP n ѐ੄ ݽഋਸ OOS ؘ੉ఠী प೷ೠ׮ 4. प೷ ౠࢿ࢚ য়ߡೖ౴ী ݒ਋ ਬ੄೧ঠೣਵ۽ प೷ਸ ݆੉ ߈ࠂೞ૑ ঋח׮ © Quant Study, Beomjun Shin, July 23, 2016
39. അप੄ प೷ ࢸ҅: ೟ण੉ ੓ח ݽ؛ 1. ୭Ӕ N֙ਸ प೷ ؘ੉ఠীࢲ ߓઁ೧ك׮(Holdout data;Test Data) 2. Walk Forward Tes9ng ೐ۨ੐(ৈ۞ߥ੄ IS/OSS ؘ੉ఠܳ ঳਺)ਵ۽ పझ౟ܳ ߈ࠂೠ׮ • Walk Forward Tes9ng ب ߈ࠂೞ׮ࠁݶ ѾҴ OSS ؘ੉ఠী য়ߡೖ౴ػ׮ • ಌನݢझܳ ҙ଴ೠ പࣻܳ ߈٘द ӝ۾ೠ׮ • प೷੄ ߈ࠂ਷ ઱۽ ݽ؛੄ ߸ച৬ ೞ੉ಌ౵ۄ޷ఠ੄ ߸ച۽ ҳࢿػ׮ 3. о੢ જও؍ TOP n ѐ੄ ݽഋਸ OOS ؘ੉ఠী प೷ೠ׮ © Quant Study, Beomjun Shin, July 23, 2016
40. ୭Ҋ੄ प೷ Real Money Trading! © Quant Study, Beomjun Shin, July 23, 2016
41. Common Evalua+on Metrics • t-sta&s&c; p-value • sharpe ra&o • Annualized Return, Annual Vola)lity • CAGR • turnover • maxdrawdown © Quant Study, Beomjun Shin, July 23, 2016
42. Sharpe Ra)o details Risk-Free Rate (Variance is not affected by constant) Benchmark © Quant Study, Beomjun Shin, July 23, 2016
43. Annualize Sharpe Ra/o © Quant Study, Beomjun Shin, July 23, 2016
44. Annualize Sharpe Ra/o • IID Return Assump.on(NO SERIAL COLLREATION! REAL?) • Daily -> Yearly: • Monthly -> Yearly: © Quant Study, Beomjun Shin, July 23, 2016
45. Addi$onal evalua$on metrics(TODO..җઁ?) • sor%no ra%o • calmar ra%o • traynor ra%o • informa%on ra%o • winning ra%o • RoMAD • beta, alpha © Quant Study, Beomjun Shin, July 23, 2016
46. All that Gli)ers Is Not Gold: Sharpe Ra5o Specifically, we find that commonly reported backtest evalua9on metrics like the Sharpe ra)o offer li.le value in predic)ng out of sample performance (R² < 0.025). In contrast, higher order moments, like vola)lity and maximum drawdown, as well as porEolio construc9on features, like hedging, show significant predic9ve value of relevance to quan9ta9ve finance prac99oners. - All that GliKers Is Not Gold. Quantopian © Quant Study, Beomjun Shin, July 23, 2016
47. All that Gli)ers Is Not Gold: Vola3lity Moreover, risk metrics that aim to quan3fy vola%lity alone like annual vola3lity (Pearson R² = 0.67; p < 0.0001), and maximum drawdown (Pearson R² = 0.34; p < 0.0001) had sta3s3cally significant correla3ons between their IS and OOS period - All that GliRers Is Not Gold. Quantopian © Quant Study, Beomjun Shin, July 23, 2016
48. ਬ੄޷ೠ ߸ࣻ ଺ӝ: 5࠙ਤ ಂఠ ࠙ࢳ & ߮஖݃௼ ࢸ੿ © Quant Study, Beomjun Shin, July 23, 2016
49. ੸੺ೠ ߮஖݃௼ ࢸ੿ • Stock Picking • ߮஖݃௼ب ݽഋҗ э਷ ܻߖ۠य ઱ӝ۽ प೷ • ࠁా਷ Equal Weighted Index • Single Asset • Buy and hold strategy ߮஖݃௼ܳ ੉ਊೞৈ ݽ؛੄ ୡҗࣻ੊ܫ ҅࢑ೠ׮ © Quant Study, Beomjun Shin, July 23, 2016
50. ߸ࣻ੄ ߸ജ [߸ࣻ]ܳ [઱ӝ]۽ [߸ഋ]೧ ࠁҊ [ҙ੼]ীࢲ [৘ஏӝр]݅ఀ ࠁਬ • ߸ࣻ: Value, Momentum, Quality • ઱ӝ: 1ѐਘ, 1௪ఠ(=3ѐਘ), 1֙, 3~5֙ • ߸ഋ: пઙ ੉زಣӐ, ݽݭథ, Valua9on Model • ҙ੼: Loser-Follow(contrarian), Winner-Follow • ৘ஏӝр: 1ѐਘ, 1௪ఠ(=3ѐਘ), 1֙, 3~5֙ © Quant Study, Beomjun Shin, July 23, 2016
51. Remind! Common Mistakes © Quant Study, Beomjun Shin, July 23, 2016
52. 1. Survivor-biased Data(Data Snooping) • NaN ؘ੉ఠ • ޷ܻ Ҋ੿ػ ਬפߡझ ೧Ѿ଼ • җѢ द੼ীࢲ੄ ਬפߡझ ؘ੉ఠ ࢎਊ © Quant Study, Beomjun Shin, July 23, 2016
53. 2. Lookahead Bias • ؘ੉ఠ݃੉׬ ݽ؛ীࢲ पࣻೞӝ ए਍ Overlap Bias • য׬ ߊ಴৬ ઙо੄ ҙ҅ (য׬਷ ઙоо աৡ ٍী ߊ಴ؽ) • ղо ݅ٚ ݽ؛੉ ইޖܻ ѐ౸੉ۄب 2008֙ ܻझ௼݅ ೖೡ ࣻ ੓׮ݶ જ਷ ੹ۚ ੉ ؽ ೧Ѿ଼ • ؘ੉ఠ੄ द੼ ੿ഛೞѱ ౵ঈೞӝ(ؘ੉ఠ ಿ౹੉!) • "য়ט੄ ઙо"ܳ NaNਵ۽ فҊ प೷ೞӝ © Quant Study, Beomjun Shin, July 23, 2016
54. 3. In-sample Backtes1ng • э਷ ؘ੉ఠܳ فҊ ৈ۞ߣ ߈ࠂ प೷ റ э਷ ؘ੉ఠীࢲ పझ౟ ೧Ѿ଼ • Mul%ple Hypothesis Tes%ng য়ܨ ੿ഛೞѱ ੉೧ೞӝ • प೷ പࣻ/प೷ ӝ۾ ԝԝೞѱ ۽Ӧ೧فӝ • IS/OOS పझ౟ ೞӝ © Quant Study, Beomjun Shin, July 23, 2016
55. 4. Market Impact, Slippage • ઙо Ѣېա زदഐо Ѣېীࢲ ղо ֍਷ ઱ޙ੉ оѺਸ ৢܽ׮ • Ѣې۝੉ ੘਷ ઙݾਸ ਬפߡझ۽ فҊ ݅ٚ ੹ۚ • ੸੺ೠ Rule of Thumb਷ ҃೷ਵ۽ ٜ݅੗ ೧Ѿ଼ • ੸੺ೠ ठܻೖ૑ܳ хউ೧ࢲ ߔపझ౴ ࣻ੊ܫ ҅࢑(e.g. S&P500਷ 5bp, ੘਷ ઙ ݾ਷ 50bp) • Ѣې۝ ؘ੉ఠ ࢎਊೞӝ(Ѣې۝ ֫਷ ઙݾ ࢎਊ; ࠁా ૑ࣻ੄ ҃਋ 10র ੉࢚?) © Quant Study, Beomjun Shin, July 23, 2016
56. 5. Overfi)ng Model • ߈ࠂ੸ਵ۽ in-sample ؘ੉ఠܳ فҊ "tweak" ೞҊ "refine"ೠ ݽ؛ • ߸ࣻо ݆Ҋ ࠂ੟ೠ ݽ؛ ࢎਊೞӝ ೧Ѿ଼ • IS/OOS పझ౟ ೞӝ • п ਬפߡझ ҳࢿ ੗࢑੉ ನ౟ಫܻয় ࣻ੊ܫী ޷஘ ੉੊ ࠺઺ ࠁӝ © Quant Study, Beomjun Shin, July 23, 2016
57. 6. Trus(ng stateful strategy luck • Rebalancing ઱ӝী ٮۄ ݽ؛ ࢿמ਷ ௼ѱ௼ѱ ߄Ո׮(੉۠ ݽ؛ਸ stateful strategyۄ ೠ ׮) • Walk-forward Tes;ngীࢲب start pointܳ য٣۽ ೞջী ٮۄ ࢿמ੉ ׳ۄ૓׮ ೧Ѿ଼ • оמೠ ݽٚ ܻߖ۠य ઱ӝ, द੘੼ ੹ࠗ ׮ܰѱ فҊ प೷(௏٘ܳ ੜ ૞ىঠೠ׮!)ೞҊ ࣻ੊ܫ ੄ Varianceܳ ࢓ಝࠁӝ • গୡী Varianceо ௼׮ݶ ੉ ੹ۚ੉ ޖ੄޷ೞ׮Ҋ ࢤп೧ঠೠ׮ • ٮۄࢲ, ୭؀ೠ 'daily' ؘ੉ఠܳ ࢎਊೞӝ © Quant Study, Beomjun Shin, July 23, 2016
58. 7. Procedure Overfi.ng • ইޖܻ ਤ੄ पٜࣻਸ ઑबೞ؊ۄب ੉޷ җѢח җѢ੉޲۽ ੺ର੸ੋ য় ߡೖ౴਷ ߈٘द ߊࢤ ೧Ѿ଼ • ݒੌ੄ पઁ ؘ੉ఠ۽ Paper Trading Nѐਘ ੉࢚ ೧ࠄ ٍ पઁ تਸ ֍ӝ • ٮۄࢲ, ୭؀ೠ ૑ࣘ੸ੋ दझమ੉ غب۾ ݒੌ ؘ੉ఠо ୶оؼ ࣻ ੓ѱ दझమ ҳࢿೞӝ © Quant Study, Beomjun Shin, July 23, 2016
59. 8. Random Strategy • ਋ো൤ աৡ ੹ۚੌө? ইקө? • ࢎਊೠ ؘ੉ఠ ԝԝೞѱ ࢓ಝࠁӝ • খ੄ 1~7ߣ੄ पࣻח ੄ب੸ਵ۽ ೞח पࣻо ইפۄ ؘ੉ఠܳ ԝԝೞѱ ࢓ೖ ૑ ঋইࢲ ߊࢤೞח पࣻ ೧Ѿ଼ • p-value ҅࢑ ١ ా҅੸ ӝߨ ࢎਊೞҊ ୭ࣗೠ੄ प೷ਸ ৈ۞ߣ ೧ࠁӝ • ݽٚ Input ؘ੉ఠী ؀ೠ द҅ৌ ࠁӝ, NaN ؘ੉ఠ ਬޖ ࠁӝ © Quant Study, Beomjun Shin, July 23, 2016
60. 9. Don't Forget Commission, Tax • ௫౟੹ۚ਷ ӝࠄ੸ਵ۽ ݒݒഥ੹ਯ੉ ֫਺. ࣻࣻܐח ߈٘द хউ೧ঠೣ ೧Ѿ଼ • ࣻࣻܐ ݽ؛੉ ٜযр ߔపझ౴ ҅࢑ ۽૒ ٜ݅যفӝ • ࣻࣻܐ ҅࢑੉ য۵׮ݶ ୭ࣗೠ ఢয়ߡח ҅࢑ೞӝ © Quant Study, Beomjun Shin, July 23, 2016
61. We need our powerful backtest tool ! • ࣚए਍ ࢎਊ & ࡅܲ ࢎਊ • ઱ਃ evalua)on metric ҅࢑(return, turnover etc.) • п ઙݾ੄ ੉੊ ҳࢿ ࠺ਯ • ࣻࣻܐ Ҋ۰ ࣻ੊ܫ ҅࢑ • пઙ Visualiza)on © Quant Study, Beomjun Shin, July 23, 2016
62. Quant ױ࢚ • ־ҳա ࠁח ؘ੉ఠ۽ ੹ۚ ٜ݅ӝ • Contrarian vs Follower, ѾҴ਷ Behavior ৘ஏऱ਑? • Sinificanceܳ ঳ӝূ ցޖաب ੸਷ ؘ੉ఠ • যڌѱ ա݅੄ ؘ੉ఠܳ ݅٘חо? © Quant Study, Beomjun Shin, July 23, 2016
63. Future Works • झझ۽ ٣పੌ ଻਋ӝ • ೧৻ ૐӂࢎ ܻನ౟ झఠ٣ • ݒ઱ HIT ֫ও؍ SSRN ಕ੉ಌ ࣁ޷ա • पઁ ై੗ೞӝ © Quant Study, Beomjun Shin, July 23, 2016
64. References • wiki: Sensi+ve and specficity • The p value and the base rate fallacy • Evalua+on Metrics • All that Gli@ers Is Not Gold • 9 Mistakes Quants Make that Cause Backtests to Lie by Tucker Balch, Ph.D © Quant Study, Beomjun Shin, July 23, 2016
65. Conven&onal Assump&ons © Quant Study, Beomjun Shin, July 23, 2016