High-Performance Factor Evaluation with DolphinDB

This factor evaluation framework uses the CSV files provided in the factor_backtest_data directory. You can download the factor_backtest_data archive from the end of this document and extract it to <YOUR_DOLPHINDB_PATH>/server/. The data can then be loaded directly with the loadText function for subsequent computation and analysis.

factorTB = loadText("<YOUR_DOLPHINDB_PATH>/server/factor_backtest_data/factorTB.csv")
mktmvTB = loadText("<YOUR_DOLPHINDB_PATH>/server/factor_backtest_data/mktmvTB.csv")
industryTB = loadText("<YOUR_DOLPHINDB_PATH>/server/factor_backtest_data/industryTB.csv")
bp_factorTB = loadText("<YOUR_DOLPHINDB_PATH>/server/factor_backtest_data/bp_factorTB.csv")
bpTB = loadText("<YOUR_DOLPHINDB_PATH>/server/factor_backtest_data/bpTB.csv")
retTB = loadText("<YOUR_DOLPHINDB_PATH>/server/factor_backtest_data/retTB.csv")
benchmarkTB = loadText("<YOUR_DOLPHINDB_PATH>/server/factor_backtest_data/benchmarkTB.csv")

Use select top 100 * from factorTB to preview the factor in the factorTB table, where trade_date is the trading date, stock_code is the stock identifier, and factor is the factor name.

Use select top 100 * from mktmvTB to preview the market capitalization in mktmvTB, where mktmv represents market capitalization of each stock.

Use select top 100 * from industryTB to preview the industry classification in industryTB, where industry indicates the industry of each stock.

Use select top 100 * from bp_factorTB to preview the multi-factor data in bp_factorTB, where both factor and bp are factor variables.

Use select top 100 * from bpTB to preview the bp factor in bpTB.

Use select top 100 * from retTB to preview the stock return in retTB, where ret denotes returns.

Use select top 100 * from benchmarkTB to preview the benchmark return.

This module does not impose strict requirements on column names. You only need to ensure that the input tables follow the same schema as the examples above. For convenience, it is recommended to use consistent column names for trading dates and security identifiers across all input tables.

In terms of data frequency, the module supports input data at daily, weekly, or monthly frequencies.

factor and bp can be processed directly using the preprocess function, which performs outlier treatment, standardization, and neutralization.

Code examples and a preview of the processed results are as follows:

use factorBacktest

//data preprocessing
factorNew = preprocess(factor=factorTB,
    factorName="factor",
    mktmvTB = mktmvTB,
    industryTB = industryTB,
    dateName = "trade_date",
    securityName = "stock_code",
    mvName = "mktmv",
    indName = "industry",
    delOutlierIf = true,
    standardizeIf = true,
    neutralizeIf = true,
    delOutlierMethod = "mad",
    standardizeMethod = "rank",
    n = 3,
    modify = false ,
    startDate = 2008.01.04,
    endDate = 2017.01.20)

// Observe the result after factor preprocessing
select top 100 * from factorNew

/*
 trade_date stock_code factor
 2008.01.04 000002 -0.06393232889798993
 2008.01.04 000006 1.5546531851632837
 2008.01.04 000009-1.2096673940460276
 2008.01.04 000011 0.7557481911404614
 2008.01.04 000012 0.32144219178486666
*/

The preprocess function exposes several key parameters:

• factorName: the name of the column that contains factor values in the input factor table. In the example, the factor values in factorTB are stored in the factor column.
• mvName and indName: the column names representing market capitalization and industry classification in the mktmvTB and industryTB tables, respectively.
• dateName and securityName: the column names for the trading date and the security identifier.

After preprocessing the factor data, stocks are sorted by factor values and divided into groups, and the return of each group is calculated. The "H-L" group represents the long-short portfolio return, calculated by going long on the highest factor group and short on the lowest factor group.

The grouped returns can be observed and formatted to six decimal places. An example of the corresponding code and results is shown below.

groupRet = singleResult["groupRet"]
groupname = ["Group0" , "Group1" , "Group2" , "Group3" , "Group4" , "H-L"]
<select top 10 trade_date , round(_$$groupname , 6)  as _$$groupname from groupRet>.eval()

Once the grouped returns are calculated, their statistical significance can be tested. In multi-factor regression models, residuals often exhibit heteroskedasticity and autocorrelation. Using standard OLS to estimate the standard errors of β may lead to inaccurate results of significance testing. This module implements a Newey–West adjustment (Newey & West, 1987) to correct for heteroskedasticity and autocorrelation in the residuals.

The core Newey–West adjustment provides a consistent covariance estimator that corrects for heteroskedasticity and autocorrelation. Specifically:

• T denotes the total number of time periods,
• e_t​ is the residual at time t,
• X_i=[x_i1,x_i2,...,x_iK]′ is the i-th row of X transposed,
• L is the maximum lag used to account for autocorrelation.

The resulting S is then substituted into the OLS variance expression V​_OLS to obtain the Newey–West heteroskedasticity- and autocorrelation-consistent covariance estimate.

In this section, the neweyWestTest function is used to estimate the mean of grouped returns and output the t- and p-values adjusted via Newey–West. The underlying computation is performed by the neweyWestCorrection function.

The long-short portfolio statistics can be accessed via singleResult["retStat"]["H-L"]. The output includes the mean return (ret_mean(%)), p-value (p_value), and t-value (t_value).

// output:
// ret_mean(%): 0.465416325805543
// p_value: 0.00000568582996640643
// t-value: 4.591311581999794

Analyzing the above results, it can be seen that the mean long-short group return rate of the factor is 0.465%, the p-value is 0.0000056, and the t-value is 4.59, indicating that the mean long-short return rate is significantly different from 0.

The information coefficient (IC) measures the cross-sectional correlation between factor values and next-period stock returns, and is commonly used to evaluate a factor’s predictive power. A larger absolute IC indicates a more effective factor. A positive IC implies that higher factor values are associated with higher future returns, while a negative IC suggests the opposite.

The information ratio (IR) is defined as the ratio of the mean excess return to its standard deviation, and reflects the stability of a factor’s ability to generate alpha.

By evaluating both IC and IR together, we can assess both the predictive strength and robustness of a factor, and determine whether it is reliable for sustained investment.

singleResult["icTest"]

/*
output:
Rank_ICIR: 0.10311019138174475
Rank_IC mean: 0.05030535697624222
IC mean: 0.03846492107792371
Factor name: 'factor'
IC standard deviation: 0.09652585561925417
IR: 0.39849344852894575
Proportion of IC>0 (%): 64.57883369330453
Proportion of IC>0.03 (%): 52.48380129589633
*/

The output obtained from singleResult["icTest"] is a dictionary with eight keys. From the results, the IC mean is 0.038, which does not meet the commonly accepted threshold for strong predictive power (≥ 0.05). The IR is 0.398, falling below the stability benchmark (≥ 0.5). Overall, the factor exhibits only weak predictive ability and relatively poor stability.

In addition, the getFactorIc function can be used to plot both time-series and cumulative charts of IC and Rank_IC. The corresponding code and results are shown below.

correlation = getFactorIc(factorNew, retTB, "factor" ,"trade_date" ,"stock_code", "ret" , 2008.01.02 , 2017.01.31)
correlation['IC_cumsum'] = correlation["IC"].cumsum()
correlation['Rank_IC_cumsum'] = correlation["Rank_IC"].cumsum()

Observe the factor IC and Rank_IC results, and set the output format to six decimal places:

//Observe factor IC and Rank_IC results
groupname = ["IC" , "Rank_IC" , "IC_cumsum" , "Rank_IC_cumsum"]
<select top 10 trade_date , round(_$$groupname , 6)  as _$$groupname from correlation>.eval()

Plot using the plot function:

plot(table(correlation["IC"] ,correlation['IC_cumsum']) , correlation["trade_date"] , title = "FTime series plot and cumulative of IC for factor", extras={multiYAxes: true})

plot(table(correlation["Rank_IC"] ,correlation['Rank_IC_cumsum']) ,
              correlation["trade_date"] , title = "Factor factor Long-Short Return Curves (Market
              Cap Weighted)", extras={multiYAxes: true})

When analyzing IC time-series and cumulative charts, different types of information can be extracted. For example, frequent sign changes in IC, as shown in Figure 3-3, indicate unstable predictive performance across market regimes. The cumulative IC chart aggregates IC values over time and reflects the factor’s long-term predictive contribution. The steadily increasing cumulative IC in Figure 3-3 suggests that the factor consistently contributes positive expected returns over the long run.

In this section, we analyze the backtest results for each factor-based portfolio group and plot both the net value curves for the groups and the long-short return curve. The results of the single-group factor analysis can be obtained from singleResult["backtest"], as shown below.

backtestTB = singleResult["backtest"]
//Observe group backtest results
groupname = ["Group0","Group1","Group2","Group3","Group4","H_L","benchmark"]
<select index , round(_$$groupname , 6) as _$$groupname from backtestTB>.eval()

plot(singleResult["netValue"], singleResult["timeIndex"], singleResult["title1"])

plot(singleResult["retTable"], singleResult["timeIndex"], singleResult["title2"])

From the results, the factor demonstrates strong return differentiation across groups in the backtest. However, the relatively high volatility indicates weaker risk control. In comparison, the long-short strategy performs better, exhibiting lower volatility and higher returns.

In factor evaluation, dependent double sorts are commonly used to assess whether combining two factors leads to better performance than using either factor alone. This approach helps examine the degree of information overlap between the two factors and identify potential information again. Specifically, for two factors X and Y, assets are sequentially sorted and grouped by both factors to construct portfolios, whose performance is then analyzed.

After computing portfolio returns based on the dependent double sorts of the factor and bp factors, we further estimate the mean returns for each group and report the corresponding Newey–West–adjusted t-statistics. The relevant code and output are shown below.

dbSortMean = doubleResult["dbSortMean"]
groupname = ["Group0" , "Group1" , "Group2" , "Group3" , "Group4" , "H_L" ]
<select index , round(_$$groupname , 6) as _$$groupname from dbSortMean>.eval()

In the results table, the index column represents the bp-based groups. The Group0~Group4 columns denote the factor-based groups. The results show that:

• Across all bp groups, the H-L long-short returns are positive, with relatively high t-values, indicating that the high-factor portfolios significantly outperform the low-factor portfolios. This suggests that the factor consistently distinguishes asset returns across different bp groups.
• For each bp group, the mean returns generally exhibit an upward trend from Group0 to Group4. This suggests that higher bp groups are associated with improved portfolio performance across different factor groups, indicating a positive impact of the bp factor on returns.
• The factor shows stronger monotonicity and statistical significance among assets with medium to high bp values (e.g., Group3 and Group4), suggesting that the effectiveness of the factor is related to asset size characteristics.

The factorBacktest module also supports the computation of backtest metrics for double-sorted portfolios. These results can be obtained using doubleResult["dbSortBacktest"].

dbSortBacktest = doubleResult["dbSortBacktest"]
groupname = ["Group0" , "Group1" , "Group2" , "Group3" , "Group4" , "H_L" ]
<select index , round(_$$groupname , 6)  as _$$groupname from dbSortBacktest>.eval()

From the backtest metrics, we find that both the bp and factor values are positively associated with portfolio performance. As the bp and factor groups increase from Group0 to Group4, the Sharpe ratio and annualized returns rise steadily, while annualized volatility and maximum drawdown decline.

To mitigate the impact of cross-sectional correlation in regression residuals on standard error estimation, the module also implements the Fama–MacBeth regression based on Fama and MacBeth (1973), Risk, Return, and Equilibrium: Empirical Tests. The functionality is provided through the famaMacbethReg function.

Similar to conventional cross-sectional regression, the Fama–MacBeth approach consists of two stages.

In the first stage, a time-series regression is performed for each asset to estimate its factor exposure β_i by regressing individual asset returns on the factor returns.

In the second stage, a cross-sectional regression is conducted at each time point t, where asset returns are regressed on their estimated factor exposures β_i.

This step is repeated independently for every time point. For example, with T=500 time points, 500 cross-sectional regressions are performed, each using only the cross-sectional data at that time. The final regression coefficients are obtained by averaging the coefficients across all time points.

Since the coefficients from the second-stage regressions may exhibit heteroskedasticity and autocorrelation, Newey–West adjustments are applied to obtain consistent standard error estimates.

The Fama–MacBeth regression results can be accessed viadoubleResult["famaMacbethResult"].

doubleResult["famaMacbethResult"]

/*
 output:
Average_Obs: 1,591.1749460043197
R-Square: 0.017901474327154446
factor:
   t-statistic: 6.986433993363168
   beta: 0.0019545565844644496
   p-value: 0.00000000000988853444
   standard error: 0.0002797645531785171
bp:
  t-statistic: 3.6404282777536614
  beta: 0.001105722959290441
  p-value: 0.00030294518028139983
  standard error: 0.0003037343067702823
*/

Note that the regressionStep parameter in famaMacbethReg is a boolean vector indicating whether the two-step Fama–MacBeth procedure is applied. By default, it is set to false, meaning that factor values are treated as factor exposures (or factor exposures are directly provided), and only the second-stage cross-sectional regression is performed.

The output is returned as a dictionary containing the model’s average R2, the average number of observations, and factor-specific statistics, including t-statistic, beta (risk premium), p-value, and standard error.

From the results, we observe the following:

• The model exhibits a relatively low R2, indicating limited explanatory power, with approximately 98% of the variation in asset returns not explained by the factor and bp factors.
• The factor has a statistically significant positive effect on asset returns, with an estimated risk premium of 0.00195.
• The bp factor also shows a statistically significant positive effect on asset returns, with an estimated risk premium of 0.00110.