diff --git a/.changelog/4523.txt b/.changelog/4523.txt
new file mode 100644
index 0000000000..551115e82c
--- /dev/null
+++ b/.changelog/4523.txt
@@ -0,0 +1,3 @@
+```release-note:enhancement
+compute: added autoscaling_policy.cpu_utilization.predictive_method field to `google_compute_autoscaler` and `google_compute_region_autoscaler`
+```
diff --git a/google-beta/resource_compute_autoscaler.go b/google-beta/resource_compute_autoscaler.go
index 355eb776c7..33b1446fe6 100644
--- a/google-beta/resource_compute_autoscaler.go
+++ b/google-beta/resource_compute_autoscaler.go
@@ -113,6 +113,16 @@ scales up until it reaches the maximum number of instances you
 specified or until the average utilization reaches the target
 utilization.`,
 									},
+									"predictive_method": {
+										Type:     schema.TypeString,
+										Optional: true,
+										Description: `Indicates whether predictive autoscaling based on CPU metric is enabled. Valid values are:
+
+- NONE (default). No predictive method is used. The autoscaler scales the group to meet current demand based on real-time metrics.
+
+- OPTIMIZE_AVAILABILITY. Predictive autoscaling improves availability by monitoring daily and weekly load patterns and scaling out ahead of anticipated demand.`,
+										Default: "NONE",
+									},
 								},
 							},
 						},
@@ -973,12 +983,22 @@ func flattenComputeAutoscalerAutoscalingPolicyCpuUtilization(v interface{}, d *s
 	transformed := make(map[string]interface{})
 	transformed["target"] =
 		flattenComputeAutoscalerAutoscalingPolicyCpuUtilizationTarget(original["utilizationTarget"], d, config)
+	transformed["predictive_method"] =
+		flattenComputeAutoscalerAutoscalingPolicyCpuUtilizationPredictiveMethod(original["predictiveMethod"], d, config)
 	return []interface{}{transformed}
 }
 func flattenComputeAutoscalerAutoscalingPolicyCpuUtilizationTarget(v interface{}, d *schema.ResourceData, config *Config) interface{} {
 	return v
 }
 
+func flattenComputeAutoscalerAutoscalingPolicyCpuUtilizationPredictiveMethod(v interface{}, d *schema.ResourceData, config *Config) interface{} {
+	if v == nil || isEmptyValue(reflect.ValueOf(v)) {
+		return "NONE"
+	}
+
+	return v
+}
+
 func flattenComputeAutoscalerAutoscalingPolicyMetric(v interface{}, d *schema.ResourceData, config *Config) interface{} {
 	if v == nil {
 		return v
@@ -1372,6 +1392,13 @@ func expandComputeAutoscalerAutoscalingPolicyCpuUtilization(v interface{}, d Ter
 		transformed["utilizationTarget"] = transformedTarget
 	}
 
+	transformedPredictiveMethod, err := expandComputeAutoscalerAutoscalingPolicyCpuUtilizationPredictiveMethod(original["predictive_method"], d, config)
+	if err != nil {
+		return nil, err
+	} else if val := reflect.ValueOf(transformedPredictiveMethod); val.IsValid() && !isEmptyValue(val) {
+		transformed["predictiveMethod"] = transformedPredictiveMethod
+	}
+
 	return transformed, nil
 }
 
@@ -1379,6 +1406,10 @@ func expandComputeAutoscalerAutoscalingPolicyCpuUtilizationTarget(v interface{},
 	return v, nil
 }
 
+func expandComputeAutoscalerAutoscalingPolicyCpuUtilizationPredictiveMethod(v interface{}, d TerraformResourceData, config *Config) (interface{}, error) {
+	return v, nil
+}
+
 func expandComputeAutoscalerAutoscalingPolicyMetric(v interface{}, d TerraformResourceData, config *Config) (interface{}, error) {
 	l := v.([]interface{})
 	req := make([]interface{}, 0, len(l))
diff --git a/google-beta/resource_compute_autoscaler_test.go b/google-beta/resource_compute_autoscaler_test.go
index 4b2f02a62a..55993e96e2 100644
--- a/google-beta/resource_compute_autoscaler_test.go
+++ b/google-beta/resource_compute_autoscaler_test.go
@@ -297,6 +297,7 @@ resource "google_compute_autoscaler" "foobar" {
     cooldown_period = 60
     cpu_utilization {
       target = 0.5
+      predictive_method = "OPTIMIZE_AVAILABILITY"
     }
     scale_down_control {
       max_scaled_down_replicas {
diff --git a/google-beta/resource_compute_region_autoscaler.go b/google-beta/resource_compute_region_autoscaler.go
index c013082718..e6a53357fd 100644
--- a/google-beta/resource_compute_region_autoscaler.go
+++ b/google-beta/resource_compute_region_autoscaler.go
@@ -113,6 +113,16 @@ scales up until it reaches the maximum number of instances you
 specified or until the average utilization reaches the target
 utilization.`,
 									},
+									"predictive_method": {
+										Type:     schema.TypeString,
+										Optional: true,
+										Description: `Indicates whether predictive autoscaling based on CPU metric is enabled. Valid values are:
+
+- NONE (default). No predictive method is used. The autoscaler scales the group to meet current demand based on real-time metrics.
+
+- OPTIMIZE_AVAILABILITY. Predictive autoscaling improves availability by monitoring daily and weekly load patterns and scaling out ahead of anticipated demand.`,
+										Default: "NONE",
+									},
 								},
 							},
 						},
@@ -972,12 +982,22 @@ func flattenComputeRegionAutoscalerAutoscalingPolicyCpuUtilization(v interface{}
 	transformed := make(map[string]interface{})
 	transformed["target"] =
 		flattenComputeRegionAutoscalerAutoscalingPolicyCpuUtilizationTarget(original["utilizationTarget"], d, config)
+	transformed["predictive_method"] =
+		flattenComputeRegionAutoscalerAutoscalingPolicyCpuUtilizationPredictiveMethod(original["predictiveMethod"], d, config)
 	return []interface{}{transformed}
 }
 func flattenComputeRegionAutoscalerAutoscalingPolicyCpuUtilizationTarget(v interface{}, d *schema.ResourceData, config *Config) interface{} {
 	return v
 }
 
+func flattenComputeRegionAutoscalerAutoscalingPolicyCpuUtilizationPredictiveMethod(v interface{}, d *schema.ResourceData, config *Config) interface{} {
+	if v == nil || isEmptyValue(reflect.ValueOf(v)) {
+		return "NONE"
+	}
+
+	return v
+}
+
 func flattenComputeRegionAutoscalerAutoscalingPolicyMetric(v interface{}, d *schema.ResourceData, config *Config) interface{} {
 	if v == nil {
 		return v
@@ -1368,6 +1388,13 @@ func expandComputeRegionAutoscalerAutoscalingPolicyCpuUtilization(v interface{},
 		transformed["utilizationTarget"] = transformedTarget
 	}
 
+	transformedPredictiveMethod, err := expandComputeRegionAutoscalerAutoscalingPolicyCpuUtilizationPredictiveMethod(original["predictive_method"], d, config)
+	if err != nil {
+		return nil, err
+	} else if val := reflect.ValueOf(transformedPredictiveMethod); val.IsValid() && !isEmptyValue(val) {
+		transformed["predictiveMethod"] = transformedPredictiveMethod
+	}
+
 	return transformed, nil
 }
 
@@ -1375,6 +1402,10 @@ func expandComputeRegionAutoscalerAutoscalingPolicyCpuUtilizationTarget(v interf
 	return v, nil
 }
 
+func expandComputeRegionAutoscalerAutoscalingPolicyCpuUtilizationPredictiveMethod(v interface{}, d TerraformResourceData, config *Config) (interface{}, error) {
+	return v, nil
+}
+
 func expandComputeRegionAutoscalerAutoscalingPolicyMetric(v interface{}, d TerraformResourceData, config *Config) (interface{}, error) {
 	l := v.([]interface{})
 	req := make([]interface{}, 0, len(l))
diff --git a/google-beta/resource_compute_region_autoscaler_test.go b/google-beta/resource_compute_region_autoscaler_test.go
index 7df486026d..8ff3b7f893 100644
--- a/google-beta/resource_compute_region_autoscaler_test.go
+++ b/google-beta/resource_compute_region_autoscaler_test.go
@@ -213,6 +213,7 @@ resource "google_compute_region_autoscaler" "foobar" {
     cooldown_period = 60
     cpu_utilization {
       target = 0.5
+      predictive_method = "OPTIMIZE_AVAILABILITY"
     }
     scale_down_control {
       max_scaled_down_replicas {
diff --git a/website/docs/r/compute_autoscaler.html.markdown b/website/docs/r/compute_autoscaler.html.markdown
index 666e892a7c..f4d3bf6f55 100644
--- a/website/docs/r/compute_autoscaler.html.markdown
+++ b/website/docs/r/compute_autoscaler.html.markdown
@@ -357,6 +357,12 @@ The `cpu_utilization` block supports:
   specified or until the average utilization reaches the target
   utilization.
 
+* `predictive_method` -
+  (Optional, [Beta](https://terraform.io/docs/providers/google/guides/provider_versions.html))
+  Indicates whether predictive autoscaling based on CPU metric is enabled. Valid values are:
+  - NONE (default). No predictive method is used. The autoscaler scales the group to meet current demand based on real-time metrics.
+  - OPTIMIZE_AVAILABILITY. Predictive autoscaling improves availability by monitoring daily and weekly load patterns and scaling out ahead of anticipated demand.
+
 The `metric` block supports:
 
 * `name` -
diff --git a/website/docs/r/compute_region_autoscaler.html.markdown b/website/docs/r/compute_region_autoscaler.html.markdown
index e300b9f61f..f7c5d52fd0 100644
--- a/website/docs/r/compute_region_autoscaler.html.markdown
+++ b/website/docs/r/compute_region_autoscaler.html.markdown
@@ -269,6 +269,12 @@ The `cpu_utilization` block supports:
   specified or until the average utilization reaches the target
   utilization.
 
+* `predictive_method` -
+  (Optional, [Beta](https://terraform.io/docs/providers/google/guides/provider_versions.html))
+  Indicates whether predictive autoscaling based on CPU metric is enabled. Valid values are:
+  - NONE (default). No predictive method is used. The autoscaler scales the group to meet current demand based on real-time metrics.
+  - OPTIMIZE_AVAILABILITY. Predictive autoscaling improves availability by monitoring daily and weekly load patterns and scaling out ahead of anticipated demand.
+
 The `metric` block supports:
 
 * `name` -