diff --git a/formula/README.md b/formula/README.md
index e4e9976f78ea53c8358deb831c9dca11a0efa00b..cb5f555e24dc3aeec9c681c41e81628653813a7d 100644
--- a/formula/README.md
+++ b/formula/README.md
@@ -20,7 +20,7 @@ formula
 | 0   | MIN           | 计算AB最小值                                            | MIN(CLOSE,HIGH)                   | [√]  | [√]  |
 | 0   | MA            | 计算N周期的移动平均值, 简称均线                                  | MA(CLOSE,5)                       | [√]  | [√]  |
 | 0   | DMA           | S序列的动态移动平均, A作为平滑因子                                | DMA(CLOSE,5)                      | [X]  | [X]  |
-| 0   | EMA           | S序列N周期的指数移动平均, α=/(1+com)                          | EMA(CLOSE,5)                      | [X]  | [X]  |
+| 0   | EMA           | S序列N周期的指数移动平均, α=/(1+com)                          | EMA(CLOSE,5)                      | [√]  | [√]  |
 | 0   | SMA           | 计算N周期的简单移动平均值                                      | SMA(CLOSE,5, 1)                   | [√]  | [√]  |
 | 0   | WMA           | S序列的N周期的加权移动平均值                                    | WMA(CLOSE,5)                      | [√]  | [√]  |
 | 0   | STD           | 计算N周期内的标准差                                         | STD(CLOSE,20)                     | [√]  | [√]  |
diff --git a/formula/ema.go b/formula/ema.go
new file mode 100644
index 0000000000000000000000000000000000000000..0b844a3e473c73614e587742b322c18f82edaa42
--- /dev/null
+++ b/formula/ema.go
@@ -0,0 +1,83 @@
+package formula
+
+import (
+	"gitee.com/quant1x/pandas"
+	"gitee.com/quant1x/pandas/exception"
+	"gitee.com/quant1x/pandas/stat"
+)
+
+// EMA 指数移动平均,为了精度 S>4*N  EMA至少需要120周期
+// alpha=2/(span+1)
+// TODO:这个版本是对的, 通达信EMA居然实现了真的序列, 那为啥SMA不是呢?!
+func EMA(S pandas.Series, N any) any {
+	var X []stat.DType
+	switch v := N.(type) {
+	case int:
+		X = stat.Repeat[stat.DType](stat.DType(v), S.Len())
+	case pandas.Series:
+		vs := v.DTypes()
+		X = stat.Align(vs, stat.DTypeNaN, S.Len())
+	default:
+		panic(exception.New(1, "error window"))
+	}
+	k := X[0]
+	x := S.EWM(pandas.EW{Span: stat.DTypeNaN, Callback: func(idx int) stat.DType {
+		j := X[idx]
+		if j == 0 {
+			j = 1
+		} else {
+			k = j
+		}
+		return stat.DType(stat.DType(2) / (j + 1))
+	}, Adjust: false}).Mean().Values()
+	_ = k
+	return x
+}
+
+// EMA_v2 通达信公式管理器上提示, EMA(S, N) 相当于SMA(S, N + 1, M=2), 骗子, 根本不对
+func EMA_v2(S pandas.Series, N any) any {
+	M := 2
+	var X float32
+	switch v := N.(type) {
+	case int:
+		X = float32(v)
+	case pandas.Series:
+		vs := v.Values()
+		fs := pandas.SliceToFloat32(vs)
+		X = fs[len(fs)-1]
+	default:
+		panic(exception.New(1, "error window"))
+	}
+	x := S.EWM(pandas.EW{Alpha: float64(M) / float64(X+1), Adjust: false}).Mean().Values()
+	return x
+}
+
+// EMA_v0 仿SMA实现, 错误
+func EMA_v0(S pandas.Series, N any) any {
+	var X float32
+	switch v := N.(type) {
+	case int:
+		X = float32(v)
+	case pandas.Series:
+		vs := v.Values()
+		fs := pandas.SliceToFloat32(vs)
+		X = fs[len(fs)-1]
+	default:
+		panic(exception.New(1, "error window"))
+	}
+	x := S.EWM(pandas.EW{Span: stat.DType(X), Adjust: false}).Mean().Values()
+	return x
+}
+
+// EMA_v1 Rolling(N), 每个都取最后一个, 错误
+func EMA_v1(S pandas.Series, N any) any {
+	x := S.Rolling(N).Apply(func(S pandas.Series, N stat.DType) stat.DType {
+		r := S.EWM(pandas.EW{Span: N, Adjust: false}).Mean().DTypes()
+		if len(r) == 0 {
+			return stat.DTypeNaN
+		}
+		return r[len(r)-1]
+	}).Values()
+	return x
+
+}
diff --git a/formula/ema_test.go b/formula/ema_test.go
new file mode 100644
index 0000000000000000000000000000000000000000..2c06b410ccd0cfccc265a3528ec3a3a24041ff74
--- /dev/null
+++ b/formula/ema_test.go
@@ -0,0 +1,45 @@
+package formula
+
+import (
+	"fmt"
+	"gitee.com/quant1x/pandas"
+	"gitee.com/quant1x/pandas/stat"
+	"github.com/viterin/vek/vek32"
+	"testing"
+)
+
+func TestEMA(t *testing.T) {
+	f0 := []float64{1, 2, 3, 4, 5, 6, 7, 8, 9}
+	s0 := pandas.NewSeriesWithoutType("f0", f0)
+	fmt.Println(EMA(s0, 7))
+	csv := "~/.quant1x/data/cn/002528.csv"
+	df := pandas.ReadCSV(csv)
+	df.SetNames("data", "open", "close", "high", "low", "volume", "amount", "zf", "zdf", "zde", "hsl")
+	fmt.Println(df)
+	//df.SetName("收盘", "close")
+	CLOSE := df.Col("close")
+
+	cs := CLOSE.Values().([]float32)
+	REF10 := REF(CLOSE, 10).([]float32)
+	v1 := vek32.Div(cs, REF10)
+	df01 := pandas.NewSeries(pandas.SERIES_TYPE_FLOAT32, "x", v1)
+	x0 := make([]stat.DType, CLOSE.Len())
+	df01.Apply(func(idx int, v any) {
+		f := v.(float32)
+		t := stat.DType(0)
+		if f >= 1.05 {
+			t = stat.DType(1)
+		}
+		x0[idx] = t
+	})
+	//x := stat.Where(v2, as, bs)
+	n := BARSLAST(pandas.NewSeries(pandas.SERIES_TYPE_FLOAT32, "", x0))
+	fmt.Println(n[len(n)-10:])
+	x := EMA(CLOSE, pandas.NewSeries(pandas.SERIES_TYPE_DTYPE, "", n))
+
+	//x := EMA(CLOSE, 7)
+	sx := pandas.NewSeries(pandas.SERIES_TYPE_DTYPE, "x", x)
+	df = pandas.NewDataFrame(CLOSE, sx)
+	fmt.Println(df)
+
+}
diff --git a/formula/sma.go b/formula/sma.go
index 485317ebbc4695eb5c288dbc6dbc99b9dc98d346..a836711840b0b93c52aac854656785fb98993876 100644
--- a/formula/sma.go
+++ b/formula/sma.go
@@ -56,7 +56,7 @@ func SMA_v5(S pandas.Series, N any, M int) any {
 	return x
 }
 
-// 听说SMA(S, N, 1) 其实就是MA(S,N), 试验后发现是骗子
+// SMA_v4 听说SMA(S, N, 1) 其实就是MA(S,N), 试验后发现是骗子
 func SMA_v4(S pandas.Series, N any, M int) any {
 	var X []float32
 	switch v := N.(type) {
@@ -72,7 +72,7 @@ func SMA_v4(S pandas.Series, N any, M int) any {
 	return S.Rolling(X).Mean().Values()
 }
 
-// 使用滑动窗口
+// SMA_v3 使用滑动窗口
 func SMA_v3(S pandas.Series, N any, M int) any {
 	if M == 0 {
 		M = 1
@@ -87,7 +87,7 @@ func SMA_v3(S pandas.Series, N any, M int) any {
 	return x
 }
 
-// 最原始的python写法
+// SMA_v1 最原始的python写法
 func SMA_v1(S pandas.Series, N int, M int) any {
 	if M == 0 {
 		M = 1
diff --git a/formula/sma_test.go b/formula/sma_test.go
index e373bd05c909bac1e1c025e0a1f529765b913093..0ddc743340d2d57619456c1133ab3dd50e356894 100644
--- a/formula/sma_test.go
+++ b/formula/sma_test.go
@@ -37,7 +37,7 @@ func TestSMA(t *testing.T) {
 	n := BARSLAST(pandas.NewSeries(pandas.SERIES_TYPE_FLOAT32, "", x))
 	fmt.Println(n[len(n)-10:])
 	//r1 := SMA(CLOSE, pandas.NewSeries(pandas.SERIES_TYPE_FLOAT32, "", n), 1)
-	r1 := SMA(CLOSE, 6, 1)
+	r1 := SMA(CLOSE, 7, 1)
 	s2 := pandas.NewSeries(pandas.SERIES_TYPE_FLOAT32, "sma", r1)
 	df2 := pandas.NewDataFrame(s2)
 	fmt.Println(df2)
diff --git a/generic_ewm.go b/generic_ewm.go
index 640cb8d2d6a9b9d34d06c00be26054f8888de650..3ad3e36b3b072ef5662239d2f071967677b40227 100644
--- a/generic_ewm.go
+++ b/generic_ewm.go
@@ -149,8 +149,9 @@ func (w ExponentialMovingWindow) notadjustedMean(data Series, alpha stat.DType,
 	var (
 		count  int
 		values = data.Values().([]stat.DType)
-		beta   = 1 - alpha
-		last   = values[0]
+		//values = data.DTypes() // Dtypes有复制功能
+		beta = 1 - alpha
+		last = values[0]
 	)
 	if Float64IsNaN(last) {
 		last = 0
diff --git a/stat/array.go b/stat/array.go
new file mode 100644
index 0000000000000000000000000000000000000000..a41e45059da263eb3ac1f2b4da2dd34abf2e7de8
--- /dev/null
+++ b/stat/array.go
@@ -0,0 +1,13 @@
+package stat
+
+type NDArray []DType
+
+type Frame interface {
+	Len() int
+}
+
+type Array[T Number] []T
+
+func (a Array[T]) Len() int {
+	return len(a)
+}
diff --git a/stat/array_test.go b/stat/array_test.go
new file mode 100644
index 0000000000000000000000000000000000000000..be63f3075bfe08de781e7854b47461f0890ea9dc
--- /dev/null
+++ b/stat/array_test.go
@@ -0,0 +1,13 @@
+package stat
+
+import (
+	"fmt"
+	"testing"
+)
+
+func TestArray_Len(t *testing.T) {
+	f1 := []float64{1, 2, 3, 4, 5}
+	a1 := Array[float64](f1)
+	fmt.Println(a1)
+	fmt.Println(a1.Len())
+}