Adaptive Cycle + RNN Anomaly Detection (NEW!)
================================================================================
PART 1: BEAST ANOMALY DETECTOR SERIES
================================================================================
Three related indicators, shared as research previews. All descend from
mcdon030's original "Beast Autoencoder-RNN" published in the ThinkOrSwim
community. The originals combined an autoencoder, a simplified RNN, and kNN
anomaly scoring into a single composite detector. These versions rewrite the
mathematical core, fix architectural bugs in the learning components, and add
new self-adaptive elements.
Research Preview | TradingView code is included
Results have not yet been formally tested or backtested.
================================================================================
SECTION A — BEAST MK II: ADAPTIVE CYCLE EDITION
================================================================================
Beast Mk II — Adaptive Cycle
Autoencoder anomaly detection with Ehlers adaptive cycle confirmation.
Note: This indicator is currently in research preview on ThinkOrSwim. Results
have not yet been formally tested or backtested. All observations are
qualitative.
The Idea
Most anomaly detectors trigger on a single condition — a volatility spike, a
volume outlier, an RSI extreme. The problem is that any one of those signals
fires constantly in normal markets. What's rare is when several structurally
different measures flag the same bar at the same time.
Beast Mk II monitors three independent views of the market — reconstruction
error from a rolling autoencoder, true k-nearest-neighbor distance on close,
and the direction of an Ehlers adaptive cycle — and only fires when they
agree, inside a market regime filter. The autoencoder tells you the market
is behaving unlike its recent self. The kNN tells you this bar is unusually
far from its nearest neighbors in recent history. The cycle tells you
momentum has recently turned. When all three line up under high volume or
expanded volatility, something structurally unusual is happening.
Plain English
Think of the main plot as a "strangeness ratio." It lives near zero during
normal chop and drifts upward when the market starts behaving unlike itself.
When it crosses the yellow dashed line at 1.0, the indicator is saying
"this bar doesn't fit the pattern of recent bars." When the ratio stays
above 1.0 for two or more bars and the other confirmations agree, the bar
turns red — that's the composite anomaly signal.
The cycle label below the plot shows whether Ehlers' adaptive cycle is
turning up or down, and whether a direction change happened within the last
3 bars (marked with an asterisk). A recent turn is one of the confirmation
signals that can gate an anomaly flag.
The alpha label shows how much the cycle has adapted to recent conditions.
Alpha starts at 0.20 and hill-climbs based on the cycle's own trade win
rate — stepping smaller (more responsive) when the cycle is too slow for
current price action, larger (more stable) when it's too fast. The win rate
in parentheses is the cycle's realized accuracy on its own direction-change
trades. It needs at least 5 completed trades before the adapter is trusted.
Reading the Chart
Visual Element What It Means
──────────────────────────────────────────────────────────────────────
Main plot (green) Reconstruction error ratio — below threshold, normal state
Main plot (orange) Ratio above 1.0 but signal not yet persistent
Main plot (red) Composite anomaly signal is active
Yellow dashed line Threshold at 1.0 — the line that matters
Cyan line EMA-smoothed error ratio — slower confirmation
"Anomaly" DETECTED or Normal, with color
"Regime" Current market regime (Volume / Volatility / Both / Normal)
"Persist" How many consecutive bars the signal has been active
"Ratio" Current ratio value in decimal form
"Cycle" Cycle direction; asterisk = recent turn
"Alpha" Adapted alpha value + cycle win rate + trade count
──────────────────────────────────────────────────────────────────────
Technical Architecture
Layer 1 — Autoencoder Reconstruction Error
Three features (VWAP, shares-per-tick, true range) are independently
normalized with z-score over a 5-bar window, smoothed through a Hull Moving
Average, and compressed via short-period averaging. The compressed signal
is then expanded back through a mirror decoder. The absolute difference
between normalized input and reconstructed output becomes the per-feature
reconstruction error. Final error is the arithmetic mean of the three
normalized errors — ensuring no single feature dominates by scale.
Layer 2 — True k-Nearest-Neighbor Distance
For each bar, the absolute difference |close - close[i]| is computed for i
in [1, lookback]. Five sequential fold passes extract the k=5 smallest
distances without needing sortable arrays (each pass finds the smallest
value strictly greater than the previous pass's winner). The kNN score is
the mean of those five nearest distances. Unlike the original "Beast"
implementation — which computed mean absolute deviation and called it kNN
— this is the actual algorithm.
Layer 3 — Ehlers Adaptive Cycle with Win-Rate Hill Climbing
A two-pole Ehlers cycle filter runs at a reference alpha of 0.20. Each
direction change of the cycle simulates a trade; realized PnL over each
trade is tracked. When the cycle's rolling win rate falls below 0.55 and
the cycle's speed is visibly mismatched to price (cycle movement > 2x
price movement = too fast; < 0.3x = too slow), alpha is nudged by one
step (default 0.01) in the corrective direction. Bounded to
[alphaMin, alphaMin + alphaSweepCount * alphaStep]. Minimum 5 completed
trades required before adapter activates.
Layer 4 — Regime Filters
Two independent regime gates must agree for an anomaly to fire:
- Volume Break: HMA(10) of Volume RSI(14) > 49
- Volatility Break: ATR(5) > ATR(20) × 1.2
Composite Signal
Fires when: autoencoder ratio > 1.0 threshold AND emaError > mean
AND (kNN anomaly OR recent cycle turn) AND (volume break OR volatility
break), persistent for at least 2 bars.
Parameters
Parameter Default Notes
──────────────────────────────────────────────────────────────────────
encodingSize 5 Autoencoder compression window
thresholdLookback 20 Window for dynamic threshold stats
thresholdMultiplier 1.5 Std deviations above mean to trigger
emaLength 15 EMA smoothing on error
minSignalLength 2 Required persistence in bars
kNNLookback 20 Bars searched for nearest neighbors
fallbackAlpha 0.20 Initial cycle alpha
alphaMin 0.05 Floor for adaptation
alphaStep 0.01 Adaptation step size
minTradesForAlpha 5 Trades before adapter trusts win rate
cycleTurnLookback 3 Bars within which a turn counts
──────────────────────────────────────────────────────────────────────
Intended for 5-minute charts. A warning label fires on other timeframes.
What's Good About It
Three independent views: Autoencoder, kNN, and cycle each come from
different mathematical families — reconstruction, distance, and recursive
filtering. They fail in different ways, so their agreement is meaningful.
Ratio-based plot: The main plot is error divided by threshold. Above 1.0
means "above threshold" regardless of current market volatility, so the
indicator reads the same during calm and busy sessions.
Self-adapting cycle: Alpha is not a user-tuned constant. It climbs a hill
defined by the cycle's own realized trade accuracy. The user sees where it
landed and how many trades drove the adaptation.
True kNN: The original Beast indicator's "kNN" was actually mean absolute
deviation. This version implements real k-nearest-neighbor distance via
sequential fold passes — slower but mathematically honest.
Regime-gated: Anomalies only fire during volume or volatility expansion.
Prevents flagging during flat tape.
What's Uncertain or Risky
Not yet tested: No backtesting has been performed. No win rate, no Sharpe,
no walk-forward validation. Treat this as a research tool, not a proven
strategy.
Cycle adapter is hill-climbing, not full search: ThinkScript can't run
parallel recursive simulations inside a fold, so the adapter uses a
single reference cycle and nudges alpha based on speed mismatch plus
win rate. This converges slowly. In whipsaw markets it may never
converge at all.
Small sample win rate: The cycle's win rate is measured on its own
direction changes. Early in the day or on slow symbols, the trade
count is small and the win rate statistically noisy.
Active hours assumption: The threshold multiplier increases by 0.2
during US equity regular hours (09:30 to 16:00 ET). Hardcoded. Futures,
crypto, and non-US markets will not match this schedule.
k=5 is fixed: Despite the kNNK input existing, the implementation uses
5 fold passes. Changing the input has no effect. Left in for interface
parity but worth knowing.
================================================================================
SECTION B — BEAST MK III-A: ADAPTIVE CYCLE + RNN
================================================================================
Beast Mk III-A — Adaptive Cycle + RNN
All of Mk II, plus a properly implemented recurrent neural network running
in parallel.
Note: This indicator is currently in research preview on ThinkOrSwim. Results
have not yet been formally tested or backtested. All observations are
qualitative.
The Idea
Mk II confirms anomalies through an adaptive cycle — a mathematical filter
that reacts to past price. Mk III-A adds a fourth independent view: a small
neural network that actively learns. The RNN reads four binary market
features, passes them through three hidden sigmoid nodes, and outputs a
probability that the next bar will close up. Weights update every bar via
online gradient descent — actual learning, not the frozen-random-weights
state the original Beast indicator shipped with.
The RNN contributes to anomaly detection through disagreement. When the
network is confident about direction (> 0.65 or < 0.35) but realized
price action has moved the other way over the last 3 bars, that confident-
but-wrong state is itself an anomaly signal. The market is doing something
the network's learned model didn't expect.
Plain English
Everything from Mk II still applies — same main plot, same three core
confirmations, same regime gates. Mk III-A adds two extra labels.
"RNN: 0.72 Up !" means the neural network's probability output is 0.72,
it's predicting up, and the exclamation mark indicates a disagreement
anomaly — the network thinks up, but recent bars have gone down.
"RNN Acc: 58%" is the rolling accuracy of the network's predictions over
the last 20 bars. Above 55% is green, below 45% is red. This tells you
whether to weight the RNN's contribution right now. A network running at
50% accuracy is guessing.
Reading the Chart
Same as Mk II, plus:
Visual Element What It Means
──────────────────────────────────────────────────────────────────────
"RNN" Network output probability + direction + anomaly flag
"RNN Acc" Rolling 20-bar prediction accuracy
──────────────────────────────────────────────────────────────────────
Technical Architecture
Layers 1–4 from Mk II are unchanged.
Layer 5 — Recurrent Neural Network
Architecture: 4 binary inputs → 3 sigmoid hidden nodes → 1 sigmoid output.
All connections are weighted; no bias term (to reduce parameter count).
Inputs:
- x1: 1 if close > close[1] else 0
- x2: 1 if Volume RSI > Volume RSI[1] else 0
- x3: 1 if TR > Average(TR, 14) else 0
- x4: 1 if close > VWAP else 0
Forward Pass (uses previous bar's inputs to predict current bar's direction):
- Hidden: h_j = sigmoid(Σ w_ij * x_i[1])
- Output: o = sigmoid(Σ v_j * h_j)
Training:
- Loss: L = 0.5 * (o - target)² where target = (close > close[1])
- Gradient: dL/dw_ij computed via standard backpropagation
- Update: w_ij[t] = w_ij[t-1] - eta * dw_ij[t-1]
- Learning rate eta is fixed at 0.10 (this is what Option C changes)
Weight initialization: (Random() - 0.5) * 0.5 on bar 1. Small values
centered near zero prevent early sigmoid saturation.
RNN Anomaly Condition:
Network output is "confident" when |o - 0.5| > 0.15. Realized move over
the last 3 bars is computed as the majority direction. Anomaly fires
when confident prediction disagrees with realized majority.
Composite Signal Modification:
Where Mk II required (kNN OR cycleTurn), Mk III-A requires
(kNN OR cycleTurn OR rnnAnomaly). All other conditions unchanged.
Parameters (additions beyond Mk II)
Parameter Default Notes
──────────────────────────────────────────────────────────────────────
eta 0.10 Fixed learning rate
rnnConfidenceBand 0.15 Distance from 0.5 to count as confident
rnnDisagreeLookback 3 Bars of realized data for disagreement
──────────────────────────────────────────────────────────────────────
Intended for 5-minute charts.
What's Good About It
Real learning: Unlike the original Beast indicator's RNN (which was broken —
weights never updated due to a recursive assignment bug), Mk III-A runs
genuine online gradient descent with proper backpropagation through a
three-node hidden layer.
Disagreement signal: The RNN doesn't just predict — it flags when its
prediction conflicts with realized price. This gives it a distinct role
in the composite signal, rather than duplicating what the cycle already
does.
Transparent accuracy: The 20-bar rolling accuracy label tells you whether
to trust the network right now. No hidden quality score.
Parallel architecture: The RNN runs independently of the cycle. If the
network is doing well (accuracy > 55%) and the cycle is also well-tuned
(alpha converged), they can both contribute confirmation. Neither can
veto the other.
What's Uncertain or Risky
Not yet tested: No backtesting has been performed. Everything below is
qualitative speculation.
Small network, binary inputs: Four binary inputs mean only 16 possible
input patterns. Three hidden nodes with 15 total weights. This is a tiny
network. It will learn something — but the ceiling on what it can learn
is low. Binary inputs discard magnitude; "close up by a penny" looks
identical to "close up by a dollar."
Target is noisy: Predicting next-bar direction from bar-over-bar signals
is one of the harder problems in price action. 5-minute bar direction
has a large random component. If accuracy hovers around 50% indefinitely,
that's the problem talking, not a bug.
Online learning drift: The network's weights can wander during regime
shifts, potentially producing long periods of worse-than-random
predictions before accuracy recovers. The 20-bar rolling accuracy gives
you a read on this — but it's after the fact.
Composite signal is more permissive: Because RNN anomaly is ORed into
confirmation, Mk III-A fires signals that Mk II wouldn't. If that turns
out to include more false positives than true ones, Mk II is the
cleaner tool.
Random initialization: Each time the chart reloads, weights are re-
randomized. Two identical charts opened seconds apart can produce
different signals on the same bars. This is a fundamental property of
running stochastic initialization in a declarative indicator language.
================================================================================
SECTION C — BEAST MK III-C: SELF-TUNING EDITION
================================================================================
Beast Mk III-C — Self-Tuning Edition
Mk III-A with an additional feedback loop: the RNN's learning rate adapts
from its own prediction accuracy.
Note: This indicator is currently in research preview on ThinkOrSwim. Results
have not yet been formally tested or backtested. All observations are
qualitative.
The Idea
In Mk III-A, the network's learning rate (eta) is fixed at 0.10. That's a
reasonable default but it's a compromise — too low and the network adapts
slowly to new regimes, too high and the network thrashes during noisy
periods. Mk III-C closes the loop: the network's own rolling prediction
accuracy determines how aggressively it should learn.
When accuracy is high (>= 55%), the network is doing well and should
stabilize — eta decreases, reducing weight volatility. When accuracy is
low (<= 45%), the current weights aren't working — eta increases, allowing
the network to explore faster. Between 45% and 55%, eta holds.
Two independent self-adaptation loops now run simultaneously:
- The cycle's alpha adapts from cycle trade win rate (as in Mk II)
- The RNN's eta adapts from RNN rolling prediction accuracy
Plain English
Mk III-C reads identically to Mk III-A on the chart, with one label
change. Where Mk III-A shows "RNN Acc: 58%", Mk III-C shows
"Acc: 58% / Eta: 0.08" — the current rolling accuracy and the current
adapted learning rate side by side.
During calm, predictable periods, eta will drift down. During choppy or
surprising periods, eta will climb. You can watch the network "decide"
to pay more or less attention to recent data.
Reading the Chart
Same as Mk III-A, with one label change:
Visual Element What It Means
──────────────────────────────────────────────────────────────────────
"Acc / Eta" Rolling accuracy + current adapted learning rate
──────────────────────────────────────────────────────────────────────
Technical Architecture
Layers 1–5 from Mk III-A are unchanged, except the fixed eta is replaced
with an adaptive eta.
Self-Tuning Learning Rate
The adaptation rule uses the previous bar's rolling accuracy (to break the
circular dependency between weights, predictions, and accuracy):
prevAccuracy = rnnAccuracy[1]
if bn < minBarsForEta:
eta = etaInitial
else if prevAccuracy >= etaAccHigh and eta[1] > etaMin + etaStep:
eta = eta[1] - etaStep # exploit — stabilize
else if prevAccuracy <= etaAccLow and eta[1] < etaMax - etaStep:
eta = eta[1] + etaStep # explore — adapt faster
else:
eta = eta[1]
eta clipped to [etaMin, etaMax]
Everything else — forward pass, backprop, weight update formula, anomaly
condition — is identical to Mk III-A.
Parameters (additions beyond Mk III-A)
Parameter Default Notes
──────────────────────────────────────────────────────────────────────
etaInitial 0.10 Starting learning rate
etaMin 0.01 Floor
etaMax 0.50 Ceiling
etaStep 0.01 Adjustment size per bar
rnnAccWindow 20 Window for rolling accuracy
minBarsForEta 30 Bars before adapter activates
etaAccHigh 0.55 Above this, decrease eta
etaAccLow 0.45 Below this, increase eta
──────────────────────────────────────────────────────────────────────
Intended for 5-minute charts.
What's Good About It
Adaptive learning rate: The network tunes itself based on how it's doing,
not based on a value someone picked six months ago. This is genuine
closed-loop adaptation.
Two adaptation loops, independent: The cycle adapts from its own win
rate; the RNN adapts from its own accuracy. Neither depends on the
other's success. If one adapter is working well and the other isn't,
the indicator doesn't collapse.
Visible feedback: The eta value in the label makes the adaptation
auditable. You can see whether the network is confident (low eta) or
searching (high eta) at any moment.
Exploit-explore pattern: The adapter implements a simple, principled
version of the exploit-explore tradeoff from reinforcement learning.
Good performance → settle. Bad performance → try new weight
configurations faster.
What's Uncertain or Risky
Not yet tested: No backtesting has been performed. Everything below
is qualitative.
One-bar adaptation lag: Because accuracy this bar depends on weights
this bar depends on eta this bar — the feedback loop is broken by using
the previous bar's accuracy to set this bar's eta. The indicator always
reacts one bar late. In principle this is fine; in practice it means
adaptation won't be instant.
Potential oscillation: If accuracy hovers near the 0.45–0.55 boundary,
eta will stabilize. If it oscillates across the boundary, eta will
oscillate too. Whether this damps out or amplifies depends on the
market. Keep an eye on whether eta is converging or bouncing.
Adaptation does not fix bad architecture: A 15-weight network on 4
binary inputs has a ceiling regardless of how well its learning rate
is tuned. Self-tuning makes the network reach its ceiling faster. It
does not raise the ceiling.
Compounding uncertainty: Mk III-C has two active adaptation loops (eta
and alpha), plus the RNN's weights, plus the autoencoder's response.
Debugging unexpected behavior is harder than in Mk II. When in doubt,
fall back to Mk II for a simpler read.
All Mk III-A risks still apply: Random initialization, small network
ceiling, binary input limitation, online learning drift.
================================================================================
CREDITS & LINEAGE
================================================================================
Original "Beast Autoencoder-RNN for Anomaly Detection": mcdon030 (2025),
enhanced by Grok. Shared in the ThinkOrSwim community under the original
file header. The architecture of the original — combining autoencoder
reconstruction error, an RNN, and a kNN-style distance score — is
preserved in these versions.
What these versions change:
- Fixed the RNN's frozen-weights bug (recursive assignment via CompoundValue)
- Replaced the simplified 2-input single-node RNN with 4-input 3-hidden-node
architecture and proper backpropagation
- Replaced mean-absolute-deviation-called-kNN with real kNN (k=5 smallest
distances via sequential fold passes)
- Removed regularization penalty that was subtracted from activations
(regularization belongs in the loss function, not the output)
- Removed injected random noise from the encoder (denoising autoencoders
need noise during training, not during streaming inference)
- Added the Ehlers adaptive cycle layer with hill-climbing alpha adaptation
(Mk II, III-A, III-C)
- Added the RNN disagreement anomaly mode (Mk III-A, III-C)
- Added self-tuning learning rate (Mk III-C only)
- Replaced raw reconstruction error plot with ratio-to-threshold plot
for consistent visual interpretation across market conditions
- Added Alert() call on fresh signals
All three versions are free, shared in the spirit of open research.
MarketFragments.com is building toward a community of rigorous, data-driven
traders — these tools are an invitation to collaborate. If you test any of
them and have results, we want to hear from you.
================================================================================
DISCLAIMERS
================================================================================
- All three indicators are research previews
- No backtesting has been performed
- No win rate, no Sharpe, no walk-forward validation, no statistical
significance testing
- Qualitative observations only
- Live trading conditions including slippage, fills, and timing may
produce outcomes that differ from chart appearance
- Not financial advice
- This is a research and educational tool
Attachments:
Tradingview
Beast Autoencoder + Adaptive Cycle for Anomaly Detection:
Thinkscript:
Beast Autoencoder + Adaptive Cycle for Anomaly Detection:
Beast Autoencoder + Adaptive Cycle + RNN for Anomaly Detection:
Beast Autoencoder + Adaptive Cycle + Self-Tuning RNN for Anomaly Detection: