Assignment overview

This report builds on hw7 (Fama-French 3 factor model for the 50 largest Taiwan stocks, TEJ data).

  • Window: 2018/01 to 2020/12 (36 monthly observations), stocks = 50 largest by market cap at 2020/12
  • Task 1: compute alpha and FF3_exp for each stock, pick the top 10 by each
  • Task 2: build four portfolios (alpha or FF3_exp, equal or value weights) and report how concentrated the value-weighted ones are
  • Task 3: build the 0050 benchmark from its adjusted price, plus a value-weighted proxy of the top 50
  • Task 4: compare performance versus 0050 (FF3 alpha, tracking error, information ratio), growth chart, short discussion
  • Units: all returns and factors are in percent per month
  • Data: the 0050 adjusted prices are embedded below. The stock and factor files are the same two TEJ files as hw7 (set data_dir in the first chunk).

1 Load data

1.1 Stock and factor files (same as hw7)

TEJ Pro exports UTF-16 text, so the bytes are read and converted first. The two files are found by their column names, so the file names do not matter.

unpack <- function(path) {
  if (grepl("\\.zip$", path, ignore.case = TRUE)) {
    exdir <- tempfile("tej_"); dir.create(exdir)
    unzip(path, exdir = exdir)
    path <- list.files(exdir, pattern = "\\.csv$", full.names = TRUE)[1]
  }
  path
}

read_tej_text <- function(path, n = file.size(path)) {
  raw <- readBin(path, "raw", n)
  if (length(raw) >= 2 && identical(raw[1:2], as.raw(c(0xff, 0xfe)))) {
    txt <- iconv(list(raw), from = "UTF-16LE", to = "UTF-8")
  } else {
    txt <- rawToChar(raw)
  }
  gsub("\r", "", txt, fixed = TRUE)
}

read_tej <- function(path) {
  out <- read.delim(text = read_tej_text(path), sep = "\t", quote = "",
                    comment.char = "", check.names = FALSE,
                    stringsAsFactors = FALSE, colClasses = "character",
                    na.strings = c("N.A.", "NA", ""))
  names(out)[1] <- "CO_ID"
  out
}

files <- list.files(data_dir, pattern = "\\.(csv|zip)$",
                    full.names = TRUE, ignore.case = TRUE)
files <- files[order(file.mtime(files), decreasing = TRUE)]

stock_path <- NULL
factor_path <- NULL
for (f in files) {
  p <- tryCatch(unpack(f), error = function(e) NA_character_)
  if (is.na(p)) next
  n <- min(600, file.size(p) - file.size(p) %% 2)
  head_txt <- tryCatch(read_tej_text(p, n), error = function(e) "")
  if (is.null(stock_path) && grepl("ROI_Month", head_txt, fixed = TRUE))
    stock_path <- p
  if (is.null(factor_path) && grepl("Market Risk Premium", head_txt, fixed = TRUE))
    factor_path <- p
}
if (is.null(stock_path) || is.null(factor_path))
  stop("Could not find the TEJ files in: ", data_dir,
       "\nFiles seen: ", paste(basename(files), collapse = ", "))

stocks <- read_tej(stock_path) %>%
  transmute(
    code   = sub(" .*", "", trimws(CO_ID)),
    name   = sub("^[^ ]+ +", "", trimws(CO_ID)),
    month  = as.Date(paste0(trimws(Date), "01"), format = "%Y%m%d"),
    ret    = as.numeric(ROI_Month),        # monthly return, %
    mktcap = as.numeric(`Market Cap.`)     # market cap, NTD million
  ) %>%
  filter(grepl("^[1-9][0-9]{3}$", code)) %>%   # normal stocks only (drops TDRs and ETFs)
  arrange(month, code)

ff <- read_tej(factor_path) %>%
  transmute(
    month = as.Date(paste0(trimws(Date), "01"), format = "%Y%m%d"),
    mkt   = as.numeric(`Market Risk Premium`),
    rf    = as.numeric(`Risk-free Interest Rate`) / 12,   # annual % -> monthly %
    smb   = as.numeric(`Size Factor (3 Factor)`),
    hml   = as.numeric(`Book to Market Factor`)
  ) %>%
  arrange(month)

cat("Stocks:", n_distinct(stocks$code), "| Months:", n_distinct(stocks$month),
    "| Factor months:", nrow(ff), "\n")
## Stocks: 1091 | Months: 98 | Factor months: 98

1.2 0050 adjusted price (embedded)

Monthly adjusted close (Close_Month) of 0050 from the TEJ table TSE/OTC Adjusted_Price(Monthly)-Ex_R+D, 2017/10 to 2025/12. The return is computed from the adjusted price and checked against TEJ’s ROI_Month.

etf_csv <- "Date,Close_Month,ROI_Month\n2017/10,16.11,4.9262\n2017/11,15.55,-3.4625\n2017/12,15.53,-0.1215\n2018/01,16.42,5.6934\n2018/02,15.97,-2.7171\n2018/03,16.09,0.7287\n2018/04,15.52,-3.5565\n2018/05,15.66,0.9375\n2018/06,15.8,0.8669\n2018/07,16.73,5.9034\n2018/08,17.0,1.6362\n2018/09,16.99,-0.0576\n2018/10,15.17,-10.7596\n2018/11,15.01,-1.0313\n2018/12,14.77,-1.6288\n2019/01,14.99,1.5355\n2019/02,15.66,4.4383\n2019/03,15.96,1.9316\n2019/04,16.7,4.6114\n2019/05,15.82,-5.2536\n2019/06,16.31,3.1228\n2019/07,16.84,3.2117\n2019/08,16.64,-1.1474\n2019/09,17.2,3.3599\n2019/10,18.32,6.5013\n2019/11,18.61,1.554\n2019/12,19.71,5.9564\n2020/01,18.89,-4.2293\n2020/02,18.61,-1.4454\n2020/03,15.99,-14.1003\n2020/04,17.95,12.2784\n2020/05,17.73,-1.228\n2020/06,18.88,6.4537\n2020/07,21.95,16.2803\n2020/08,21.53,-1.9268\n2020/09,21.78,1.1786\n2020/10,21.78,0.0001\n2020/11,23.96,9.9998\n2020/12,25.85,7.8992\n2021/01,27.7,7.194\n2021/02,28.83,4.0562\n2021/03,29.12,1.012\n2021/04,30.38,4.3412\n2021/05,29.63,-2.4893\n2021/06,30.03,1.3493\n2021/07,29.47,-1.8364\n2021/08,30.41,3.1604\n2021/09,29.69,-2.3512\n2021/10,29.42,-0.912\n2021/11,29.9,1.6202\n2021/12,31.52,5.435\n2022/01,31.34,-0.5322\n2022/02,30.66,-2.1548\n2022/03,30.58,-0.2888\n2022/04,28.51,-6.7705\n2022/05,28.74,0.8155\n2022/06,25.64,-10.786\n2022/07,26.54,3.5063\n2022/08,26.45,-0.3389\n2022/09,23.27,-12.0322\n2022/10,22.28,-4.2533\n2022/11,26.17,17.466\n2022/12,24.78,-5.2857\n2023/01,27.2,9.6627\n2023/02,27.32,0.4227\n2023/03,27.96,2.3569\n2023/04,27.0,-3.4539\n2023/05,29.01,7.4532\n2023/06,29.69,2.3385\n2023/07,30.05,1.2274\n2023/08,29.33,-2.4068\n2023/09,28.62,-2.4264\n2023/10,28.27,-1.223\n2023/11,30.65,8.4193\n2023/12,31.6,3.1217\n2024/01,32.08,1.5225\n2024/02,34.22,6.6617\n2024/03,37.7,10.1882\n2024/04,37.78,0.2218\n2024/05,39.64,4.8973\n2024/06,44.52,12.3193\n2024/07,43.4,-2.5107\n2024/08,43.52,0.2765\n2024/09,44.15,1.4338\n2024/10,46.03,4.2677\n2024/11,44.94,-2.3721\n2024/12,46.98,4.5393\n2025/01,48.98,4.2495\n2025/02,46.84,-4.3714\n2025/03,42.03,-10.2597\n2025/04,41.02,-2.4025\n2025/05,43.73,6.6136\n2025/06,47.07,7.616\n2025/07,50.53,7.3503\n2025/08,51.46,1.8428\n2025/09,56.65,10.0954\n2025/10,63.46,12.0245\n2025/11,61.11,-3.7066\n2025/12,64.3,5.2124"

etf <- read.csv(text = etf_csv, stringsAsFactors = FALSE) %>%
  transmute(
    month = as.Date(paste0(Date, "/01"), format = "%Y/%m/%d"),
    close = Close_Month,
    roi   = ROI_Month
  ) %>%
  arrange(month) %>%
  mutate(ret = (close / dplyr::lag(close) - 1) * 100)     # 0050 monthly return, %

cat("0050 months:", nrow(etf), "| from", format(min(etf$month)),
    "to", format(max(etf$month)), "\n")
## 0050 months: 99 | from 2017-10-01 to 2025-12-01
cat("Max |my return - TEJ ROI| (percentage points):",
    round(max(abs(etf$ret - etf$roi), na.rm = TRUE), 3), "\n")
## Max |my return - TEJ ROI| (percentage points): 0.103

2 Task 1: Alpha and FF3_exp, top 10 by each

Same first window as hw7. For each stock:

(stock return - rf) = alpha + b * MKT + s * SMB + h * HML

  • alpha = regression intercept (% per month)
  • FF3_exp = return explained by the factors = b * mean(MKT) + s * mean(SMB) + h * mean(HML), using the factor averages of the window (% per month, alpha not included)

Stocks with fewer than 36 months in the window are excluded.

formation <- as.Date("2020-12-01")
win_start <- as.Date("2018-01-01")
win_end   <- as.Date("2020-12-01")
n_months  <- length(seq(win_start, win_end, by = "month"))   # 36

members <- stocks %>%
  filter(month == formation, !is.na(mktcap)) %>%
  arrange(desc(mktcap)) %>%
  slice(1:50) %>%
  mutate(rank = row_number()) %>%
  select(rank, code, name, mktcap)

win <- stocks %>%
  filter(code %in% members$code, month >= win_start, month <= win_end) %>%
  inner_join(ff, by = "month") %>%
  mutate(ex_ret = ret - rf)

full_codes <- win %>% filter(!is.na(ret)) %>% count(code) %>%
  filter(n == n_months) %>% pull(code)

fmean <- ff %>%
  filter(month >= win_start, month <= win_end) %>%
  summarise(mkt = mean(mkt), smb = mean(smb), hml = mean(hml))

fit_one <- function(d) {
  m <- lm(ex_ret ~ mkt + smb + hml, data = d)
  b <- coef(m)
  data.frame(code    = d$code[1],
             alpha   = unname(b[1]),
             b_mkt   = unname(b[2]),
             b_smb   = unname(b[3]),
             b_hml   = unname(b[4]),
             ff3_exp = unname(b[2] * fmean$mkt + b[3] * fmean$smb + b[4] * fmean$hml))
}

win_full <- win %>% filter(code %in% full_codes)
res <- bind_rows(lapply(split(win_full, win_full$code), fit_one)) %>%
  left_join(members, by = "code") %>%
  arrange(rank)

cat("Stocks used (full 36 months):", nrow(res), "of 50\n")
## Stocks used (full 36 months): 48 of 50

Top 10 by alpha:

top_alpha <- res %>% arrange(desc(alpha)) %>% slice(1:10)
top_alpha %>%
  select(Rank = rank, Code = code, Name = name, Alpha = alpha, FF3_exp = ff3_exp) %>%
  kable(digits = 3)
Rank Code Name Alpha FF3_exp
36 2379 RT 3.213 0.981
34 6415 Silergy 3.141 1.386
33 3034 Novatek 3.049 1.112
43 2345 Accton 2.995 1.157
8 2303 UMC 2.555 1.882
3 2454 MediaTek 2.314 0.980
1 2330 TSMC 1.506 1.421
28 2327 Yageo 1.383 2.875
45 1590 Airtac 1.001 1.211
26 2395 Advantech 0.891 1.176

Top 10 by FF3_exp:

top_exp <- res %>% arrange(desc(ff3_exp)) %>% slice(1:10)
top_exp %>%
  select(Rank = rank, Code = code, Name = name, Alpha = alpha, FF3_exp = ff3_exp) %>%
  kable(digits = 3)
Rank Code Name Alpha FF3_exp
35 2603 EMC 0.382 3.479
28 2327 Yageo 1.383 2.875
32 5871 Chailease 0.554 2.185
27 2408 NTC -0.515 2.118
8 2303 UMC 2.555 1.882
44 9910 FT 0.741 1.735
19 2207 Hotai Motor 0.803 1.655
2 2317 Hon Hai -1.656 1.592
6 2308 DELTA 0.779 1.590
48 1402 FENC -0.725 1.517

3 Task 2: Four portfolios and concentration

The stocks are picked with the 2018-2020 window and held out of sample from 2021/01 to 2025/12. Portfolios are rebalanced every month:

  • Equal weight (EW): 1/N each month
  • Value weight (VW): last month’s market cap / sum of last month’s market caps of the portfolio stocks (no look-ahead)
  • If a stock has no return in a month, the weights are re-scaled over the stocks that do.
hold_start <- as.Date("2021-01-01")
hold_end   <- as.Date("2025-12-01")

panel <- stocks %>%
  filter(code %in% members$code, month >= formation, month <= hold_end) %>%
  group_by(code) %>%
  arrange(month, .by_group = TRUE) %>%
  mutate(cap_lag = dplyr::lag(mktcap)) %>%
  ungroup() %>%
  filter(month >= hold_start, !is.na(ret), !is.na(cap_lag))

build <- function(codes, scheme) {
  d <- panel %>% filter(code %in% codes)
  if (scheme == "EW") {
    d <- d %>% group_by(month) %>% mutate(w = 1 / n()) %>% ungroup()
  } else {
    d <- d %>% group_by(month) %>% mutate(w = cap_lag / sum(cap_lag)) %>% ungroup()
  }
  d
}

defs <- list(
  "Alpha-EW"  = list(top_alpha$code, "EW"),
  "Alpha-VW"  = list(top_alpha$code, "VW"),
  "FF3exp-EW" = list(top_exp$code,   "EW"),
  "FF3exp-VW" = list(top_exp$code,   "VW"),
  "Top50-VW"  = list(members$code,   "VW")     # proxy of the top 50 (Task 3)
)

wts <- bind_rows(lapply(names(defs), function(nm) {
  build(defs[[nm]][[1]], defs[[nm]][[2]]) %>% mutate(port = nm)
}))

port_ret <- wts %>%
  group_by(port, month) %>%
  summarise(ret = sum(w * ret), .groups = "drop")

3.1 How concentrated are the value-weighted portfolios?

HHI = sum of squared weights. Effective N = 1 / HHI (an equal-weight portfolio of N stocks has effective N = N).

conc <- wts %>%
  filter(grepl("VW", port)) %>%
  group_by(port, month) %>%
  summarise(max_w = max(w),
            top3  = sum(sort(w, decreasing = TRUE)[1:3], na.rm = TRUE),
            hhi   = sum(w^2), .groups = "drop") %>%
  arrange(port, month) %>%
  group_by(port) %>%
  summarise(`Largest weight, first month (%)` = 100 * first(max_w),
            `Largest weight, average (%)`     = 100 * mean(max_w),
            `Top 3 weight, average (%)`       = 100 * mean(top3),
            `HHI, average`                    = mean(hhi),
            `Effective N, average`            = mean(1 / hhi))
kable(conc, digits = 2)
port Largest weight, first month (%) Largest weight, average (%) Top 3 weight, average (%) HHI, average Effective N, average
Alpha-VW 80.66 81.92 92.08 0.68 1.48
FF3exp-VW 30.46 34.84 64.26 0.19 5.36
Top50-VW 43.64 45.94 54.57 0.22 4.69

Three biggest holdings in the first holding month (2021/01):

wts %>%
  filter(grepl("VW", port), month == hold_start) %>%
  group_by(port) %>%
  arrange(desc(w), .by_group = TRUE) %>%
  slice(1:3) %>%
  ungroup() %>%
  transmute(Portfolio = port, Code = code, Name = name, `Weight (%)` = 100 * w) %>%
  kable(digits = 1)
Portfolio Code Name Weight (%)
Alpha-VW 2330 TSMC 80.7
Alpha-VW 2454 MediaTek 7.0
Alpha-VW 2303 UMC 3.4
FF3exp-VW 2317 Hon Hai 30.5
FF3exp-VW 2308 DELTA 16.3
FF3exp-VW 2303 UMC 14.0
Top50-VW 2330 TSMC 43.6
Top50-VW 2317 Hon Hai 4.0
Top50-VW 2454 MediaTek 3.8

4 Task 3: Benchmarks

  • 0050: monthly return from its adjusted price (section 1.2).
  • Top50-VW: value-weighted proxy of the 50 largest stocks of 2020/12 (membership fixed), same monthly weighting rule as above.
rets <- port_ret %>%
  bind_rows(etf %>%
              filter(month >= hold_start, month <= hold_end) %>%
              transmute(port = "0050", month, ret)) %>%
  pivot_wider(names_from = port, values_from = ret) %>%
  arrange(month) %>%
  inner_join(ff, by = "month")

cat("Holding months:", nrow(rets), "(", format(min(rets$month)), "to",
    format(max(rets$month)), ")\n")
## Holding months: 60 ( 2021-01-01 to 2025-12-01 )

5 Task 4: Performance versus 0050

Over 2021/01 - 2025/12:

  • FF3 alpha: intercept of (portfolio - rf) on MKT, SMB, HML (% per month, with t-stat)
  • Tracking error (TE): std. dev. of (portfolio - 0050) x sqrt(12), annualized %
  • Information ratio (IR): mean active return x 12 / TE
perf <- function(r, bench, d) {
  dd <- data.frame(ex = r - d$rf, mkt = d$mkt, smb = d$smb, hml = d$hml)
  m  <- lm(ex ~ mkt + smb + hml, data = dd)
  s  <- summary(m)
  act <- r - bench
  te  <- sd(act) * sqrt(12)
  data.frame(
    `Ann. return (%)`    = (prod(1 + r / 100)^(12 / length(r)) - 1) * 100,
    `Ann. vol (%)`       = sd(r) * sqrt(12),
    `FF3 alpha (%/mo)`   = unname(coef(m)[1]),
    `t(alpha)`           = unname(s$coefficients[1, 3]),
    `Tracking error (%)` = te,
    `Info. ratio`        = mean(act) * 12 / te,
    check.names = FALSE)
}

ports <- c("Alpha-EW", "Alpha-VW", "FF3exp-EW", "FF3exp-VW", "Top50-VW", "0050")
perf_tab <- bind_rows(lapply(ports, function(p) perf(rets[[p]], rets[["0050"]], rets)))
perf_tab <- cbind(Portfolio = ports, perf_tab, stringsAsFactors = FALSE)
# 0050 against itself: TE and IR are not meaningful
perf_tab[perf_tab$Portfolio == "0050", c("Tracking error (%)", "Info. ratio")] <- NA
kable(perf_tab, digits = 3)
Portfolio Ann. return (%) Ann. vol (%) FF3 alpha (%/mo) t(alpha) Tracking error (%) Info. ratio
Alpha-EW 13.395 21.983 -0.247 -0.543 14.078 -0.373
Alpha-VW 24.430 26.105 0.289 1.112 8.312 0.614
FF3exp-EW 17.816 19.431 -0.181 -0.460 13.901 -0.137
FF3exp-VW 20.502 22.057 -0.014 -0.031 12.940 0.069
Top50-VW 20.090 18.675 0.127 1.293 2.387 -0.039
0050 19.991 19.637 0.105 0.904

5.1 Growth chart (2020/12 = 100)

growth <- sapply(ports, function(p) c(100, 100 * cumprod(1 + rets[[p]] / 100)))
dates  <- c(formation, rets$month)
cols   <- c("firebrick", "orange", "steelblue", "navy", "darkgreen", "black")

matplot(dates, growth, type = "l", lty = 1, lwd = c(2, 2, 2, 2, 2, 3), col = cols,
        xlab = "", ylab = "Value of 100", main = "Growth of 100: portfolios vs 0050")
legend("topleft", legend = ports, col = cols, lwd = 2, bty = "n", cex = 0.85)

6 Discussion

Numbers from the tables: the best annualized return among the five portfolios is Alpha-VW (24.4% a year, against 20% for 0050). The best information ratio is Alpha-VW (0.61). 0 of the 5 portfolios have an FF3 alpha with |t| above 2.

Points to comment on (edit to match your results):

  • Alpha vs FF3_exp: did picking by past alpha (stock-specific return) or by past FF3_exp (factor exposure) work better out of sample?
  • Is the alpha real? A t-stat with |t| below about 2 means the FF3 alpha is not significantly different from zero.
  • EW vs VW: value-weighted portfolios follow the biggest stocks (see the concentration table), so they stay closer to 0050 and have lower tracking error.
  • Tracking error and IR: a high TE means the portfolio moves away from 0050. The IR shows whether that extra risk was paid for.
  • Caveats: only one 36-month estimation window, 10 stocks per portfolio, picks fixed at 2020/12, and past alpha often does not persist.

7 Save outputs

write.csv(top_alpha, "hw8_top10_alpha.csv", row.names = FALSE)
write.csv(top_exp,   "hw8_top10_ff3exp.csv", row.names = FALSE)
write.csv(conc,      "hw8_concentration.csv", row.names = FALSE)
write.csv(perf_tab,  "hw8_performance.csv", row.names = FALSE)
write.csv(rets %>% mutate(month = format(month, "%Y-%m")),
          "hw8_monthly_returns.csv", row.names = FALSE)
cat("Saved 5 csv files in:", data_dir, "\n")
## Saved 5 csv files in: C:/Users/DELL/Downloads/rdataa