fix(power_added): seed pre-window miner state and apply 7-day rolling mean

Two fixes to eliminate artifacts in the Raw/QAP Power Added charts:

1. Seed-anchor: miners with power established before height 4000000 had no
   baseline, causing their full historical power to appear as a spike on day 1
   of the analysis window (+894 PiB raw). Fix injects each miner's last known
   state at height ≤ 4000000 as a virtual anchor entry the day before the
   window opens, so fill-forward starts from the correct baseline.

2. 7-day trailing rolling mean: large SP events sometimes split across day
   boundaries in power_actor_claims, creating paired +/- spikes that net to
   small real changes. Rolling mean smooths these without hiding real trends.
   Chart subtitles updated to reflect the smoothing.

Result: raw spike reduced from ±894/±238 PiB → ±41/±49 PiB; outlier days
(|raw| > 20 PiB) reduced from 148 → 45.

Author: Tanisha-fil

datasets/power_added.py

@@ -19,8 +19,14 @@
 -- power_actor_claims is change-driven: a miner only has a row on days their power changed.
 -- To get stable total-network-power per day we must fill-forward each miner's last known
 -- power to days they are absent, then sum across all miners, then diff day-over-day.
+--
+-- Bug fix: miners whose power was established before height 4000000 had no baseline,
+-- so their full historical power appeared as a spike on day 1 of the analysis window.
+-- We fix this by seeding each miner's initial state from their last known power at or
+-- before height 4000000 and injecting it as a virtual anchor entry on the day before
+-- the analysis window opens.
 with activity as (
-    -- Last known power per miner per day (only days they had activity)
+    -- Last known power per miner per day (only days they had activity in-window)
     select
         datetime_trunc(timestamp_seconds(((height * 30) + 1598306400)), day) as date,
         miner_id,
@@ -40,68 +46,99 @@
     from activity
     where rn = 1
 ),
-date_spine  as (select distinct date    from daily_activity),
-miner_bounds as (
-    select miner_id, min(date) as first_seen_date
-    from daily_activity
-    group by 1
+-- Seed: last known power per miner at or before the window boundary.
+-- Only fetch miners that actually appear in the analysis window (inner join)
+-- to keep the cross-join bounded.
+seed_raw as (
+    select
+        p.miner_id,
+        cast(p.raw_byte_power    as float64) as raw_byte_power,
+        cast(p.quality_adj_power as float64) as quality_adj_power,
+        row_number() over (partition by p.miner_id order by p.height desc) as rn
+    from `lily-data.lily.power_actor_claims` p
+    inner join (select distinct miner_id from daily_activity) m using (miner_id)
+    where p.height <= 4000000
+),
+seed_latest as (
+    select miner_id, raw_byte_power, quality_adj_power
+    from seed_raw
+    where rn = 1
+),
+-- Virtual anchor row: inject seed power as an entry on the day before the window opens.
+-- The fill-forward carries these values into day 1, so the first real diff is correct.
+seed_anchor as (
+    select
+        date_sub((select min(date) from daily_activity), interval 1 day) as date,
+        s.miner_id,
+        s.raw_byte_power,
+        s.quality_adj_power
+    from seed_latest s
+),
+-- Merge the anchor entries with actual in-window activity
+all_activity as (
+    select date, miner_id, raw_byte_power, quality_adj_power from daily_activity
+    union all
+    select date, miner_id, raw_byte_power, quality_adj_power from seed_anchor
 ),
--- All date × miner combinations from each miner's first-seen day onward
+date_spine  as (select distinct date    from all_activity),
+miner_spine as (select distinct miner_id from all_activity),
+-- All date × miner combinations (includes the anchor date)
 all_combos as (
     select d.date, m.miner_id
-    from date_spine d
-    join miner_bounds m
-      on d.date >= m.first_seen_date
+    from date_spine d cross join miner_spine m
 ),
 -- Fill-forward: carry each miner's last known power through subsequent days
 filled as (
     select
         a.date,
         a.miner_id,
-        last_value(da.raw_byte_power    ignore nulls) over (
+        last_value(aa.raw_byte_power    ignore nulls) over (
             partition by a.miner_id order by a.date
             rows between unbounded preceding and current row
         ) as raw_byte_power,
-        last_value(da.quality_adj_power ignore nulls) over (
+        last_value(aa.quality_adj_power ignore nulls) over (
             partition by a.miner_id order by a.date
             rows between unbounded preceding and current row
         ) as quality_adj_power
     from all_combos a
-    left join daily_activity da on a.date = da.date and a.miner_id = da.miner_id
-),
--- First observation per miner has no prior baseline, so skip it in delta sums.
-deltas as (
-    select
-        date,
-        miner_id,
-        case
-            when lag(raw_byte_power) over (partition by miner_id order by date) is null then null
-            else raw_byte_power - lag(raw_byte_power) over (partition by miner_id order by date)
-        end as raw_delta,
-        case
-            when lag(quality_adj_power) over (partition by miner_id order by date) is null then null
-            else quality_adj_power - lag(quality_adj_power) over (partition by miner_id order by date)
-        end as qap_delta
-    from filled
+    left join all_activity aa on a.date = aa.date and a.miner_id = aa.miner_id
 ),
 network_daily as (
     select
         date,
-        sum(coalesce(raw_delta, 0)) as raw_power_added_bytes,
-        sum(coalesce(qap_delta, 0)) as qap_power_added_bytes
-    from deltas
+        sum(coalesce(raw_byte_power,    0)) as total_raw,
+        sum(coalesce(quality_adj_power, 0)) as total_qap
+    from filled
     group by 1
 )
 select
     date,
-    raw_power_added_bytes / pow(1024, 5) as raw_power_added,
-    qap_power_added_bytes / pow(1024, 5) as quality_adjusted_power_added
+    (total_raw - lag(total_raw) over (order by date)) / pow(1024, 5) as raw_power_added,
+    (total_qap - lag(total_qap) over (order by date)) / pow(1024, 5) as quality_adjusted_power_added
 from network_daily
+-- Exclude the synthetic anchor day; it exists only to seed the fill-forward
+where date >= (select min(date) from daily_activity)
 order by date desc
 """
 
 data = client.query(sql).to_arrow(create_bqstorage_client=False)
 
-df = pl.DataFrame(data).with_columns(pl.col("date").dt.strftime("%Y-%m-%d"))
+# Apply 7-day trailing rolling mean to smooth single-day spikes caused by
+# timing/reporting artifacts in power_actor_claims (e.g. large power changes
+# split across day boundaries). The chart subtitle reflects this smoothing.
+df = (
+    pl.DataFrame(data)
+    .with_columns(pl.col("date").dt.strftime("%Y-%m-%d"))
+    .sort("date")
+    .with_columns([
+        pl.col("raw_power_added")
+            .rolling_mean(window_size=7, min_samples=1)
+            .alias("raw_power_added"),
+        pl.col("quality_adjusted_power_added")
+            .rolling_mean(window_size=7, min_samples=1)
+            .alias("quality_adjusted_power_added"),
+    ])
+    .sort("date", descending=True)
+)
 
 df.write_json(f"public/{Path(__file__).stem}.json")

src/pages/index.astro

@@ -295,14 +295,16 @@ const renewalRateSubtitleSuffix =
 					data={powerAdded}
 					x="date"
 					y="raw_power_added"
-					label="Raw Power Added"
+					label="Raw Power Added (7-day avg)"
+					subtitle="7-day rolling mean of daily raw byte power delta. Source: power_actor_claims"
 					yLabel="PiB"
 				/>
 				<LineChart
 					data={powerAdded}
 					x="date"
 					y="quality_adjusted_power_added"
-					label="Quality Adjusted Power Added"
+					label="Quality Adjusted Power Added (7-day avg)"
+					subtitle="7-day rolling mean of daily QAP delta. Source: power_actor_claims"
 					yLabel="PiB (QAP)"
 				/>
 				<LineChart