diff --git a/promql/bench.go b/promql/bench.go
new file mode 100644
index 000000000..a60091e31
--- /dev/null
+++ b/promql/bench.go
@@ -0,0 +1,45 @@
+// Copyright 2015 The Prometheus Authors
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, softwar
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package promql
+
+import "testing"
+
+// A Benchmark holds context for running a unit test as a benchmark.
+type Benchmark struct {
+	b         *testing.B
+	t         *Test
+	iterCount int
+}
+
+// NewBenchmark returns an initialized empty Benchmark.
+func NewBenchmark(b *testing.B, input string) *Benchmark {
+	t, err := NewTest(b, input)
+	if err != nil {
+		b.Fatalf("Unable to run benchmark: %s", err)
+	}
+	return &Benchmark{
+		b: b,
+		t: t,
+	}
+}
+
+// Run runs the benchmark.
+func (b *Benchmark) Run() {
+	b.b.ReportAllocs()
+	b.b.ResetTimer()
+	for i := 0; i < b.b.N; i++ {
+		b.t.RunAsBenchmark(b)
+		b.iterCount++
+	}
+}
diff --git a/promql/functions.go b/promql/functions.go
index 433c06f02..9de0ba98e 100644
--- a/promql/functions.go
+++ b/promql/functions.go
@@ -184,6 +184,102 @@ func funcIrate(ev *evaluator, args Expressions) model.Value {
 	return resultVector
 }
 
+// Calculate the trend value at the given index i in raw data d.
+// This is somewhat analogous to the slope of the trend at the given index.
+// The argument "s" is the set of computed smoothed values.
+// The argument "b" is the set of computed trend factors.
+// The argument "d" is the set of raw input values.
+func calcTrendValue(i int, sf, tf float64, s, b, d []float64) float64 {
+	if i == 0 {
+		return b[0]
+	}
+
+	x := tf * (s[i] - s[i-1])
+	y := (1 - tf) * b[i-1]
+
+	// Cache the computed value.
+	b[i] = x + y
+
+	return b[i]
+}
+
+// Holt-Winters is similar to a weighted moving average, where historical data has exponentially less influence on the current data.
+// Holt-Winter also accounts for trends in data. The smoothing factor (0 < sf < 1) effects how historical data will effect the current
+// data. A lower smoothing factor increases the influence of historical data. The trend factor (0 < tf < 1) effects
+// how trends in historical data will effect the current data. A higher trend factor increases the influence.
+// of trends. Algorithm taken from https://en.wikipedia.org/wiki/Exponential_smoothing titled: "Double exponential smoothing".
+func funcHoltWinters(ev *evaluator, args Expressions) model.Value {
+	mat := ev.evalMatrix(args[0])
+
+	// The smoothing factor argument.
+	sf := ev.evalFloat(args[1])
+
+	// The trend factor argument.
+	tf := ev.evalFloat(args[2])
+
+	// Sanity check the input.
+	if sf <= 0 || sf >= 1 {
+		ev.errorf("invalid smoothing factor. Expected: 0 < sf < 1 got: %f", sf)
+	}
+	if tf <= 0 || tf >= 1 {
+		ev.errorf("invalid trend factor. Expected: 0 < tf < 1 got: %f", sf)
+	}
+
+	// Make an output vector large enough to hold the entire result.
+	resultVector := make(vector, 0, len(mat))
+
+	// Create scratch values.
+	var s, b, d []float64
+
+	var l int
+	for _, samples := range mat {
+		l = len(samples.Values)
+
+		// Can't do the smoothing operation with less than two points.
+		if l < 2 {
+			continue
+		}
+
+		// Resize scratch values.
+		if l != len(s) {
+			s = make([]float64, l)
+			b = make([]float64, l)
+			d = make([]float64, l)
+		}
+
+		// Fill in the d values with the raw values from the input.
+		for i, v := range samples.Values {
+			d[i] = float64(v.Value)
+		}
+
+		// Set initial values.
+		s[0] = d[0]
+		b[0] = d[1] - d[0]
+
+		// Run the smoothing operation.
+		var x, y float64
+		for i := 1; i < len(d); i++ {
+
+			// Scale the raw value against the smoothing factor.
+			x = sf * d[i]
+
+			// Scale the last smoothed value with the trend at this point.
+			y = (1 - sf) * (s[i-1] + calcTrendValue(i-1, sf, tf, s, b, d))
+
+			s[i] = x + y
+		}
+
+		samples.Metric.Del(model.MetricNameLabel)
+		resultVector = append(resultVector, &sample{
+			Metric:    samples.Metric,
+			Value:     model.SampleValue(s[len(s)-1]), // The last value in the vector is the smoothed result.
+			Timestamp: ev.Timestamp,
+		})
+	}
+
+	return resultVector
+}
+
 // === sort(node model.ValVector) Vector ===
 func funcSort(ev *evaluator, args Expressions) model.Value {
 	// NaN should sort to the bottom, so take descending sort with NaN first and
@@ -848,6 +944,12 @@ var functions = map[string]*Function{
 		ReturnType: model.ValVector,
 		Call:       funcHistogramQuantile,
 	},
+	"holt_winters": {
+		Name:       "holt_winters",
+		ArgTypes:   []model.ValueType{model.ValMatrix, model.ValScalar, model.ValScalar},
+		ReturnType: model.ValVector,
+		Call:       funcHoltWinters,
+	},
 	"irate": {
 		Name:       "irate",
 		ArgTypes:   []model.ValueType{model.ValMatrix},
diff --git a/promql/functions_test.go b/promql/functions_test.go
new file mode 100644
index 000000000..bbb51df37
--- /dev/null
+++ b/promql/functions_test.go
@@ -0,0 +1,73 @@
+// Copyright 2015 The Prometheus Authors
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package promql
+
+import "testing"
+
+func BenchmarkHoltWinters4Week5Min(b *testing.B) {
+	input := `
+clear
+load 5m
+    http_requests{path="/foo"}    0+10x8064
+
+eval instant at 4w holt_winters(http_requests[4w], 0.3, 0.3)
+    {path="/foo"} 20160
+`
+
+	bench := NewBenchmark(b, input)
+	bench.Run()
+
+}
+
+func BenchmarkHoltWinters1Week5Min(b *testing.B) {
+	input := `
+clear
+load 5m
+    http_requests{path="/foo"}    0+10x2016
+
+eval instant at 1w holt_winters(http_requests[1w], 0.3, 0.3)
+    {path="/foo"} 20160
+`
+
+	bench := NewBenchmark(b, input)
+	bench.Run()
+}
+
+func BenchmarkHoltWinters1Day1Min(b *testing.B) {
+	input := `
+clear
+load 1m
+    http_requests{path="/foo"}    0+10x1440
+
+eval instant at 1d holt_winters(http_requests[1d], 0.3, 0.3)
+    {path="/foo"} 20160
+`
+
+	bench := NewBenchmark(b, input)
+	bench.Run()
+}
+
+func BenchmarkChanges1Day1Min(b *testing.B) {
+	input := `
+clear
+load 1m
+    http_requests{path="/foo"}    0+10x1440
+
+eval instant at 1d changes(http_requests[1d])
+    {path="/foo"} 20160
+`
+
+	bench := NewBenchmark(b, input)
+	bench.Run()
+}
diff --git a/promql/test.go b/promql/test.go
index 2ca105efa..60acb9285 100644
--- a/promql/test.go
+++ b/promql/test.go
@@ -411,6 +411,32 @@ func (cmd clearCmd) String() string {
 	return "clear"
 }
 
+// RunAsBenchmark runs the test in benchmark mode.
+// This will not count any loads or non eval functions.
+func (t *Test) RunAsBenchmark(b *Benchmark) error {
+	for _, cmd := range t.cmds {
+
+		switch cmd.(type) {
+		// Only time the "eval" command.
+		case *evalCmd:
+			err := t.exec(cmd)
+			if err != nil {
+				return err
+			}
+		default:
+			if b.iterCount == 0 {
+				b.b.StopTimer()
+				err := t.exec(cmd)
+				if err != nil {
+					return err
+				}
+				b.b.StartTimer()
+			}
+		}
+	}
+	return nil
+}
+
 // Run executes the command sequence of the test. Until the maximum error number
 // is reached, evaluation errors do not terminate execution.
 func (t *Test) Run() error {
diff --git a/promql/testdata/functions.test b/promql/testdata/functions.test
index 28233033e..a568595a7 100644
--- a/promql/testdata/functions.test
+++ b/promql/testdata/functions.test
@@ -288,3 +288,33 @@ eval_ordered instant at 50m sort_desc(http_requests)
 	http_requests{group="production", instance="1", job="api-server"} 200
 	http_requests{group="production", instance="0", job="api-server"} 100
 	http_requests{group="canary", instance="2", job="api-server"} NaN
+
+# Tests for holt_winters
+clear
+
+# positive trends
+load 10s
+	http_requests{job="api-server", instance="0", group="production"}	0+10x1000 100+30x1000
+	http_requests{job="api-server", instance="1", group="production"}	0+20x1000 200+30x1000
+	http_requests{job="api-server", instance="0", group="canary"}		0+30x1000 300+80x1000
+	http_requests{job="api-server", instance="1", group="canary"}		0+40x2000
+
+eval instant at 8000s holt_winters(http_requests[1m], 0.01, 0.1)
+	{job="api-server", instance="0", group="production"} 8000
+	{job="api-server", instance="1", group="production"} 16000
+	{job="api-server", instance="0", group="canary"} 24000
+	{job="api-server", instance="1", group="canary"} 32000
+
+# negative trends
+clear
+load 10s
+	http_requests{job="api-server", instance="0", group="production"}	8000-10x1000
+	http_requests{job="api-server", instance="1", group="production"}	0-20x1000
+	http_requests{job="api-server", instance="0", group="canary"}		0+30x1000 300-80x1000
+	http_requests{job="api-server", instance="1", group="canary"}		0-40x1000 0+40x1000
+
+eval instant at 8000s holt_winters(http_requests[1m], 0.01, 0.1)
+	{job="api-server", instance="0", group="production"} 0
+	{job="api-server", instance="1", group="production"} -16000
+	{job="api-server", instance="0", group="canary"} 24000
+	{job="api-server", instance="1", group="canary"} -32000