Empirical Study on Behavioral Factors in the Taiwan Stock Market: An Analysis of Broker Branch Trading Factors (Part 1)

1. Introduction: Transitioning from Chip Myths to Rational Quantification—Does Following Branch Order Flow Generate Alpha?

The Taiwan stock market is renowned for its high retail investor participation and exceptional market transparency. Every day after the market close, the Taiwan Stock Exchange (TWSE) and the Taipei Exchange (TPEx) publish detailed  Broker Branch Trading Database. This dataset itemizes the exact buy and sell volume and transaction values executed by every brokerage branch across all listed equities. Compared with centralized depository shareholder distribution reports published weekly, brokerage branch data offers daily frequency, clear identification of market entities, and trade directionality. These features make it the premier tool for investors analyzing market participant positioning.

However, many investors who trade based on branch flow fall into the trap of buying at market tops. Behavioral finance demonstrates that information asymmetries and psychological biases directly manifest in trading behaviors. On one hand, institutional investors and large accounts with genuine information advantages (smart money) execute sizable, concentrated orders, driving subsequent price momentum. On the other hand, when a stock gains widespread retail attention, excessive retail buying often drives crowd sentiment and asset congestion, causing prices to temporarily detach from fundamental value before undergoing mean reversion.

Rather than relying on intuition, data provides definitive insights

Utilizing the TEJ  Broker Branch Trading Database, this study decomposes raw trade logs into 20 behavioral factors across cash flow dynamics, major player positioning, retail activity, and crowd behavior. Through multi-year cross-sectional backtests, we evaluate whether chip-based stock selection generates sustainable alpha in the Taiwan stock market and identify the key predictive signals for live implementation.

➡️Explore Broker Trading Database!

2. Data and Behavioral Factors

This section outlines the dataset, sample selection criteria, universe construction methodology, and mathematical definitions for the 20 brokerage branch behavioral factors.

2.1 Data Sources and Setup

All data is sourced from the TEJ Database, with standardized processing applied to ensure cross-year consistency.

• Data Sources: TEJ  Broker Branch Trading Database (TWN/ABSR) and TEJ Equity Price Database (TWN/APRCD).
• Sample Scope: Taiwan Stock Exchange (TWSE) and Taipei Exchange (TPEx) common stocks, averaging approximately 1,595 equities per weekly universe.
• Universe Screening Criteria: Exclude stocks listed for less than one year, the bottom 5% by market capitalization, the bottom 5% by cumulative trailing 12-month turnover value, and penny stocks trading below NT$10.
• Backtest Period: December 2020 through June 2026 (245 weeks).
• Rebalancing Frequency: Factors are sampled on the final trading day of each week. Positions are entered at the next trading day (t+1) close and held for 5 trading days.
• Portfolio Weighting: Equal weighting.
• Execution Setup: Signal generation uses post-market branch data. Execution occurs at the t+1 close to eliminate look-ahead bias.

2.2 Factor Definitions and Formulations

Let B_i,j,t and S_i,j,t denote the buy and sell share volumes of stock i by branch j on day t, with BA_i,j,t and SA_i,j,t representing corresponding transaction amounts, and NS_i,j,t = B_i,j,t – S_i,j,t representing net buy volume. N_i,t denotes the total number of active trading branches. MA_k(·) and Skew_k(·) denote k-day moving averages and moving skewness. Factors outside the cash flow category require at least 30 active trading branches on day t to filter out illiquidity distortions.

(1) Branch Cash Flow Dynamics

These factors measure shifts in average order size per branch and branch participation relative to historical baselines. Informed traders execute concentrated, large orders to capture information rents, making abnormal order sizes indicative of informed capital flow. Expanding branch participation reflects elevated market attention. Average transaction, inflow, and outflow amounts are calculated across all branches, buying branches, and selling branches, and evaluated against a 200-day baseline.

(2) Branch Participation Force

This factor class measures buy-sell pressure through trader consensus rather than dollar volume. Let N_i,t^B = |B_i,t| and N_i,t^S = |S_i,t| represent buying and selling branch counts, while N_i,t⁺ and N_i,t^– denote net buyer and net seller counts. When buyer counts substantially exceed seller counts, it indicates sentiment-driven chasing rather than informed demand, implying negative forward returns. Skewness metrics capture right-tail spikes in buyer concentration, implying positive forward returns.

(3) Retail Branch Behavior

Retail branches are defined as branches with average order sizes under 5,000 shares per trade on day t, requiring at least 30 active retail branches. Let N_i,t^R+ and N_i,t^R- represent net buying and net selling retail branches. Herding among retail traders reflects overoptimism rather than fundamental insights, yielding negative forward return expectations.

(4) Major Player Behavior

Major player branches are defined daily as the top 30 branches ranked by absolute net transaction volume |NS_i,j,t|. Top 30 net buyer and net seller branches define sets T_i,t⁺ and T_i,t^–. Following smart money hypotheses, major player net positioning exhibits positive predictive value. Major player turnover measures liquidity absorption relative to total shares outstanding.

(5) Branch Herding Effects

Modifying the LSV herding metric, branch herding evaluates directional consensus by comparing the proportion of net-buying branches for stock i against the cross-sectional market mean p_t. Retail herding replaces total branch counts with retail branch counts.

(6) Branch Concentration and Dispersal

These factors measure buy/sell concentration and major player cost dispersion. Major player divergence measures the coefficient of variation of average purchase prices across the top 30 net buyers. High dispersion reflects fragmented valuations across entry points. Relative seller Gini coefficients standardize sell volume concentration against 100-day baselines; seller concentration typically reflects one-off liquidation, implying positive expected returns. Major player Herfindahl-Hirschman Index (HHI) differentials subtract top 30 net seller concentration from net buyer concentration; buyer overconcentration reflects narrow market breadth, implying negative expected returns.

Table 2.1 Definition of Brokerage Branch Behavioral Factors

Factor Category & Name	Code	Computation Summary	Expected Direction
Branch Cash Flow
Average Transaction Change	brkamtchg	Change in total average transaction size vs. 200-day baseline	Positive
Average Inflow Change	brkinchg	Change in average buy order size vs. 200-day baseline	Positive
Average Outflow Change	brkoutchg	Change in average sell order size vs. 200-day baseline	Positive
Branch Participation Change	brkpartchg	Change in active trading branch count vs. 100-day baseline	Positive
Branch Participation Force
Branch Order Count Power	brkcntpow	(Buyers – Sellers) / Total active branches, 10-day MA	Negative
Net Branch Ratio	brkcntratio	Net buyer branches / Net seller branches, 10-day MA	Negative
Order Count Power Skewness	brkcntskw	100-day rolling skewness of branch order count power	Positive
Retail Branch Behavior
Retail Order Count Power	brkrcntpow	Retail (Net buyers – Net sellers) / Total retail, 10-day MA	Negative
Retail Share Volume Power	brkrvolpow	Retail (Buy shares – Sell shares) / Total shares, 10-day MA	Negative
Retail Amount Power	brkramtpow	Retail (Buy amount – Sell amount) / Total amount, 10-day MA	Negative
Major Player Behavior
Major Share Volume Power	brkmvolpow	Top 30 major (Buy – Sell shares) / Total shares, 10-day MA	Positive
Major Amount Power	brkmamtpow	Top 30 major (Buy – Sell amount) / Total amount, 10-day MA	Positive
Major Buy Crowding	brkmbuycrd	Top 30 net buy volume / 100-day MA volume, 10-day MA	Positive
Major Sell Crowding	brkmsellcrd	Top 30 net sell volume / 100-day MA volume, 10-day MA	Negative
Major Turnover Rate	brkmturnrat	Top 30 major volume / Shares outstanding, 10-day MA	Positive
Branch Herding Effects
Retail Herding Index	brkrherdidx	Retail net buyer ratio – Market average ratio, 10-day MA	Negative
Overall Herding Index	brkherdidx	Total net buyer ratio – Market average ratio, 10-day MA	Negative
Concentration & Dispersal
Major Player Divergence	brkmdivg	CV of purchase price for top 30 net buyers, 10-day MA	Positive
Relative Seller Gini	brkginirel	Seller volume Gini coefficient / 100-day MA	Positive
Major HHI Differential	brkmhhidiff	Top 30 net buy HHI – Top 30 net sell HHI, 10-day MA	Negative

3. Cross-Sectional Predictive Power of Behavioral Factors

We evaluate cross-sectional return predictability using Fama-MacBeth regressions and decile portfolio sorts to measure factor information content prior to transaction cost analysis.

3.1 Fama-MacBeth Regressions

This subsection uses FM regressions to examine the cross-sectional predictive power of behavioral factors on stock returns.

Table 3.1 presents the FM regression results for each behavioral factor. Empirical results show that the average regression coefficients across factors range between -0.0011 and 0.0010, meaning that for every one cross-sectional standard deviation increase in factor value, next-week stock returns change by approximately -0.11% to 0.10%. Among the 20 factors, 15 factors are statistically significant at least at the 10% significance level (with 14 reaching the 5% level), and all factor coefficient signs align with expected directions listed in the summary table.

Regarding positive predictive factors, both major player buy/sell share power and amount power show positive correlation at the 1% significance level, supporting the smart money hypothesis that directional net positions of major players hold an information advantage. Rates of change in average branch inflow/outflow amounts and participating branch counts are also significantly positive, indicating that expanding order sizes and participation breadth imply increases in informed capital and market attention.

Regarding negative predictive factors, their coefficients are all negatively correlated at the 1% significance level, showing that “crowding” characteristics—such as overwhelming buyer counts, aligned retail net buying, and excessive buyer concentration—all signal subsequent return reversals.

On the other hand, Retail Share Volume Power, Retail Amount Power, Major Buy Crowding, Major Sell Crowding, and Major Player Divergence do not reach statistical significance at the 10% level after controlling for one-month momentum, market capitalization, and book-to-market ratios. Their predictive information may have been absorbed by the control variables, or their relationship with forward stock returns may be non-linear. We further examine these dynamics using portfolio sorts in the next subsection.

Table 3.1 Fama-MacBeth Regression Results for Behavioral Factors

Factor Category & Name	Intercept	Factor Slope	1M Momentum	Market Cap	Book-to-Market
Branch Cash Flow
Average Transaction Change	0.0035* (1.95)	0.0005* (1.83)	0.0009** (1.96)	0.0001 (0.21)	0.0000 (0.01)
Average Inflow Change	0.0035* (1.95)	0.0006** (2.15)	0.0009* (1.88)	0.0001 (0.24)	0.0000 (0.04)
Average Outflow Change	0.0035* (1.95)	0.0007** (2.51)	0.0008* (1.89)	0.0000 (0.16)	0.0000 (0.03)
Branch Participation Change	0.0035* (1.94)	0.0005** (2.05)	0.0009** (2.00)	0.0001 (0.41)	-0.0000 (-0.02)
Branch Participation Force
Branch Order Count Power	0.0036* (1.95)	-0.0011*** (-3.85)	0.0007 (1.56)	-0.0001 (-0.25)	-0.0001 (-0.20)
Net Branch Ratio	0.0036* (1.95)	-0.0006*** (-3.24)	0.0009* (1.91)	0.0001 (0.32)	-0.0001 (-0.24)
Order Count Power Skewness	0.0037** (1.96)	0.0009*** (5.74)	0.0011** (2.35)	-0.0001 (-0.24)	-0.0002 (-0.44)
Retail Branch Behavior
Retail Order Count Power	0.0036* (1.94)	-0.0010*** (-4.55)	0.0006 (1.38)	-0.0000 (-0.10)	-0.0001 (-0.26)
Retail Share Volume Power	0.0036* (1.94)	-0.0001 (-0.62)	0.0010** (2.16)	-0.0001 (-0.27)	-0.0001 (-0.33)
Retail Amount Power	0.0036* (1.94)	-0.0001 (-0.64)	0.0010** (2.15)	-0.0001 (-0.26)	-0.0001 (-0.33)
Major Player Behavior
Major Share Volume Power	0.0036* (1.95)	0.0009*** (4.73)	0.0008 (1.57)	-0.0000 (-0.14)	-0.0002 (-0.47)
Major Amount Power	0.0036* (1.95)	0.0009*** (4.72)	0.0008 (1.61)	-0.0000 (-0.13)	-0.0002 (-0.46)
Major Buy Crowding	0.0037** (1.96)	0.0006 (1.43)	0.0009* (1.67)	0.0000 (0.08)	-0.0002 (-0.55)
Major Sell Crowding	0.0037** (1.96)	-0.0004 (-1.03)	0.0010* (1.92)	-0.0000 (-0.02)	-0.0002 (-0.51)
Major Turnover Rate	0.0037* (1.95)	0.0010** (1.99)	0.0008** (1.99)	-0.0001 (-0.36)	0.0001 (0.45)
Branch Herding Effects
Retail Herding Index	0.0036* (1.94)	-0.0010*** (-4.55)	0.0006 (1.38)	-0.0000 (-0.10)	-0.0001 (-0.26)
Overall Herding Index	0.0036* (1.95)	-0.0009*** (-4.17)	0.0007 (1.56)	-0.0000 (-0.06)	-0.0001 (-0.27)
Concentration & Dispersal
Major Player Divergence	0.0036* (1.95)	0.0008 (1.51)	0.0008** (2.13)	0.0002 (0.58)	0.0001 (0.24)
Relative Seller Gini	0.0037* (1.85)	0.0006*** (3.81)	0.0011** (2.50)	0.0001 (0.15)	-0.0002 (-0.39)
Major HHI Differential	0.0036* (1.94)	-0.0008*** (-6.74)	0.0012*** (2.63)	-0.0000 (-0.02)	-0.0001 (-0.26)

3.2 Portfolio Sorts

Compared to Fama–MacBeth regressions, which rely on parametric linearity assumptions, portfolio sorts capture non-linear relationships between factor values and forward stock returns. Specifically, stocks are sorted weekly into equal-weighted decile portfolios (Q1 to Q10) based on factor ranks and held for a one-week rebalancing horizon. Return monotonicity from Q1 (lowest) to Q10 (highest) and High-minus-Low (H-L) long-short spread performances are evaluated to assess forward return predictability.

Table 3.2 presents decile portfolio performance from October 2021 through June 2026. Across all 20 behavioral factors, H-L long-short weekly spreads remain statistically distinguishable from zero at the 10% significance level or better.

For positive predictive factors, decile portfolio returns increase monotonically with factor ranks for Major Amount Power, Major Share Volume Power, Order Count Power Skewness, Branch Participation Change, Relative Seller Gini, Average Outflow Change, and Average Transaction Change. Major Amount Power generates weekly average returns scaling from 0.16% in Q1 to 0.64% in Q10, producing a long-short H-L spread of 0.47% (t=6.07), the highest t-statistic across all evaluated factors. Major Share Volume Power delivers an identical 0.47% weekly spread (t=6.04). Major Turnover Rate achieves the highest absolute long-short return at 0.58% (t=2.57), though higher return volatility results in a relatively lower t-statistic. These findings confirm that directional accumulation and liquidity turnover by major player branches convey persistent positive information advantages in Taiwan equities.

For negative predictive factors, decile portfolio returns exhibit monotonic decline across Retail Order Count Power, Retail Herding Index, Overall Herding Index, Branch Order Count Power, and Major HHI Differential, with corresponding H-L weekly spreads of -0.50% (t=-6.05), -0.50% (t=-6.05), -0.48% (t=-5.76), -0.49% (t=-5.74), and -0.23% (t=-5.61), all significant at the 1% level. The first four factors reflect buyer consensus; synchronized net buying by high volumes of branches (particularly retail accounts) precedes price mean reversion, validating the role of retail investors as uninformed noise traders.

Significantly, Retail Share Volume Power, Retail Amount Power, Major Buy Crowding, and Major Player Divergence show insignificant linear slopes in cross-sectional regressions, yet achieve long-short H-L spread significance at the 5% level under portfolio sorts. This demonstrates that their return predictability operates primarily through non-linear tail effects concentrated in extreme quantiles or overlaps with control variables like short-term momentum.

Table 3.2 Decile Portfolio Weekly Returns for Behavioral Factors (%)

Factor Category & Name	Q1	Q2	Q3	Q4	Q5	Q6	Q7	Q8	Q9	Q10	H-L Spread
Branch Cash Flow
Average Transaction Change	0.19	0.16	0.23	0.26*	0.31*	0.32**	0.38**	0.47***	0.52***	0.64***	0.45*** (3.35)
Average Inflow Change	0.19	0.16	0.24	0.26	0.30*	0.35**	0.35**	0.46***	0.55***	0.61***	0.42*** (3.33)
Average Outflow Change	0.17	0.20	0.26	0.27*	0.28*	0.35**	0.39**	0.43**	0.50***	0.61***	0.44*** (3.39)
Branch Participation Change	0.16	0.25	0.29*	0.33*	0.36**	0.37**	0.39**	0.35**	0.43**	0.54***	0.38*** (3.69)
Branch Participation Force
Branch Order Count Power	0.61***	0.58***	0.46**	0.38*	0.41**	0.34*	0.29	0.29*	0.19	0.12	-0.49*** (-5.74)
Net Branch Ratio	0.58***	0.53***	0.37*	0.40**	0.38*	0.32*	0.32*	0.32*	0.25	0.20	-0.38*** (-5.04)
Order Count Power Skewness	0.29	0.24	0.29*	0.31*	0.28*	0.36**	0.37**	0.50***	0.46**	0.59***	0.30*** (4.49)
Retail Branch Behavior
Retail Order Count Power	0.63***	0.52***	0.42**	0.43**	0.37*	0.31	0.35*	0.27	0.24	0.13	-0.50*** (-6.05)
Retail Share Volume Power	0.42***	0.48***	0.37**	0.39**	0.35*	0.37*	0.35*	0.36*	0.35*	0.24	-0.18*** (-3.04)
Retail Amount Power	0.43***	0.47***	0.37*	0.40**	0.34*	0.37*	0.34*	0.36*	0.35**	0.24	-0.19*** (-3.16)
Major Player Behavior
Major Share Volume Power	0.16	0.29	0.26	0.34*	0.40**	0.34**	0.37**	0.41**	0.46**	0.63***	0.47*** (6.04)
Major Amount Power	0.16	0.30	0.27	0.34*	0.38**	0.37**	0.37**	0.39**	0.46**	0.64***	0.47*** (6.07)
Major Buy Crowding	0.28	0.34	0.33	0.36*	0.37**	0.38**	0.33**	0.31**	0.39***	0.60***	0.32** (2.20)
Major Sell Crowding	0.55***	0.35**	0.36**	0.31**	0.36**	0.37**	0.37**	0.34*	0.37	0.30	-0.25* (-1.71)
Major Turnover Rate	0.15	0.22*	0.29**	0.30*	0.34**	0.36*	0.39*	0.35	0.53**	0.73**	0.58** (2.57)
Branch Herding Effects
Retail Herding Index	0.63***	0.52***	0.42**	0.43**	0.37*	0.31	0.35*	0.27	0.24	0.13	-0.50*** (-6.05)
Overall Herding Index	0.60***	0.53***	0.42**	0.41**	0.37*	0.32*	0.36*	0.28	0.26	0.12	-0.48*** (-5.76)
Concentration & Dispersal
Major Player Divergence	0.18*	0.24*	0.30*	0.31*	0.34*	0.36*	0.42**	0.46**	0.52**	0.55**	0.37** (2.06)
Relative Seller Gini	0.30*	0.32*	0.29	0.34*	0.33*	0.47**	0.38*	0.41**	0.43**	0.53***	0.23*** (3.50)
Major HHI Differential	0.38***	0.34**	0.46**	0.34*	0.39*	0.39*	0.39*	0.42**	0.43**	0.15	-0.23*** (-5.61)

Major Amount Power demonstrates optimal overall predictive performance, delivering a Fama-MacBeth slope significant at the 1% level (t=4.72), the highest portfolio spread t-statistic (t=6.07), and strong return monotonicity across deciles (rank correlation of 0.95).

Part 2 evaluates signal implementability, portfolio turnover, and net-of-fee performance using this representative factor!