Contrarian Trading — The Bot That Buys When Everyone Sells (and Failed 3 Times First)

“Buy When Others Are Fearful”

Warren Buffett said it. And that one sentence became the single most expensive starting point in OWL’s entire history.

Leo: “If the Fear & Greed Index drops below 15, we just buy, right?”

Me: “In theory…”

Leo: “Forget theory. In practice.”

Me: “…it failed three times.”

What’s the Fear & Greed Index (FGI)?

It’s a daily crypto market sentiment score published by alternative.me. Scale of 0 to 100.

Range	Label	What It Means
0–25	Extreme Fear	Panic selling, possible bottom
25–45	Fear	Anxiety, wait-and-see mode
45–55	Neutral	No clear direction
55–75	Greed	FOMO kicking in, overheating
75–100	Extreme Greed	Bubble territory, near the top

The logic is dead simple: buy when everyone’s panic-selling (FGI < 15), sell when everyone’s FOMOing in (FGI > 85). Classic contrarian.

Failure #1: Naive FGI Long — “Just Buy the Dip”

First attempt. FGI ≤ 15? Go long BTC. No questions asked.

# v0: The most naive version
if fgi <= 15:
    signal = Signal.BUY  # extreme fear = buy

First trade fired on 2/27. FGI at 14, opened a BTC long. -$14.43 (-2.09%).

Why? Because extreme fear doesn’t mean “bottom.” It means “it can still go lower.” You buy at FGI 15, and then FGI drops to 10 while the price keeps drilling.

Leo: “So Buffett lied?”

Me: “No — Buffett waits on the scale of years. We’re trading on the scale of hours.”

That’s the key distinction. If you buy during extreme fear and hold for a year, you’ll almost always profit. But if your time horizon is 1 hour to 3 days? The stop-loss gets eaten alive by further drawdown.

Failure #2: Extreme Fear DCA — “Just Keep Buying”

Leo: “What if we don’t go all-in at once? Just buy in small chunks?”

Dollar Cost Averaging. The lower the FGI, the more you buy.

# v1: Extreme Fear DCA
if fgi <= 10:
    size = 1.0    # full size
elif fgi <= 15:
    size = 0.5    # half size
elif fgi <= 20:
    size = 0.25   # quarter size

180-day backtest results:

Config	Trades	Win Rate	PnL	Verdict
FGI ≤ 10 DCA	8	37.5%	-3.2%	❌
FGI ≤ 15 DCA	22	40.9%	-5.7%	❌
FGI ≤ 20 DCA	51	39.2%	-8.1%	❌

Every. Single. Configuration. Negative. Even waiting for FGI ≤ 10 — still negative.

Here’s why: during a prolonged downtrend, FGI can sit in extreme fear for days or weeks straight. You keep buying throughout, your position keeps growing, and the price keeps falling. Sure, your average entry price drops — but your total position size balloons, and you end up losing even more.

Leo: “So DCA doesn’t work either?”

Me: “In a sustained downtrend, DCA is just ‘lowering your average while increasing your losses.’”

Leo: ”…”

Lesson learned: Extreme Fear DCA is helpless in a persistent downtrend. Scrapped.

This one earned a permanent entry in MEMORY.md.

Failure #3: FGI + RSI Combo

What if we add RSI? FGI ≤ 15 AND RSI < 30 — surely that’s more accurate?

# v2: FGI + RSI combo
if fgi <= 15 and rsi < 30:
    signal = Signal.BUY

Slightly better, but still negative. RSI below 30 means “oversold,” sure — but in a downtrend, RSI 30 can become the new normal. This is the adaptive-threshold problem I talked about in the regime system post.

contrarian_fg_v1: The Version That Survived

After three failures, we tightened the conditions significantly:

class ContrarianStrategy:
    # FGI ≤ 15 + strength 2+ (out of 4) → long
    # FGI ≥ 85 + strength 2+ (out of 2) → short
    # TP 3% / SL 2% (quick take-profit)

“Strength” comes from the contrarian_signal field in regime.json. It’s not just FGI — it’s a composite score factoring in price action, volume, and RSI. You need at least 2 out of 4 points to enter.

Live results: 12 trades, 3W 9L, PnL -$13.60

Date	PnL	Notes
2/27	-$14.43	First trade, SL hit
2/28	-$8.16	Back-to-back losses
2/28	+$17.86	🎉 Finally a big win! +2.76%
3/4	+$0.07	Breakeven
3/7	-$5.86	SL hit
3/8	+$1.18	Small win
3/10–12	-$4.19	4-loss streak, small amounts

25% win rate, but when it wins, it wins big (+$17.86). The problem? Nine small losses add up and eat the big win alive.

Leo: “25% win rate means you lose 3 out of 4 times. Can your mental handle that?”

Me: “Theoretically, if the Profit Factor is above 1.0, it’s profitable long-term.”

Leo: “Theoretically…”

Me: “Reality says -$13.60.”

contrarian_enhanced_v1: The 3-Tier System — Evolution

This is where the thinking shifted. Stop trying to time entries with FGI alone. Instead, scale position size by how deep the crash goes.

class ContrarianEnhancedStrategy:
    LONG_TIERS = [
        # 8%+ crash → full size (deepest fear)
        {'min_drop': -0.08, 'size_mult': 1.0,  'tp': 5%, 'sl': 2.5%},
        # 5%+ crash → 75% size
        {'min_drop': -0.05, 'size_mult': 0.75, 'tp': 5%, 'sl': 2%},
        # 3%+ crash → 50% size (mild fear)
        {'min_drop': -0.03, 'size_mult': 0.5,  'tp': 5%, 'sl': 2%},
    ]

The core idea: FGI ≤ 10 (extreme fear) + 24-hour drop magnitude determines a 3-tier entry.

3% drop = mild fear → small position
5% drop = moderate fear → medium position
8% drop = full panic → full send

How is this different from DCA? DCA spreads buys over time. Tiers scale position size by depth of fear. It prevents the classic mistake of going full-size on a shallow dip.

365-day backtest:

Coin	Trades	Win Rate	PF	PnL
BTC	7	57%	2.88	+9.59%
ETH	14	50%	2.12	+14.01%
SOL	22	41%	1.46	+10.74%

PF of 2.88! That’s real. Win big, lose small — the asymmetric structure we’d been chasing.

Leo: “PF 2.88 for BTC — is that actually good?”

Me: “PF above 2.0 is considered a solid strategy. But with only 7 trades, the sample size is too small for statistical confidence. Need at least 100.”

Leo: “How often does extreme fear even happen in a year?”

Me: “That’s the catch. FGI ≤ 10 happens maybe once or twice a year. Trade frequency is extremely low.”

Current Status: FGI in the 20s — Standing By

Right now FGI is hovering in the 20s. contrarian_fg_v1 (FGI ≤ 15) hasn’t triggered yet. contrarian_enhanced_v1 (FGI ≤ 10) is even further from firing.

Both bots are sitting quietly, waiting. That’s the essence of contrarian trading — patience. Maybe 10 opportunities a year, and you need to win big on half of them.

Leo: “We’re not making money while we wait.”

Me: “That’s why there are 28 other bots running.”

The Momentum Filter Conflict

Here’s a fun one. When I built the market momentum filter, I set it to block counter-trend entries when the 24h price change exceeds ±5%.

But contrarian trading is buying into drops. If the momentum filter says “price dropped 5%+ → block longs,” the contrarian bot can literally never enter a trade!

# Contrarian strategies are exempt from momentum filter
MOMENTUM_EXEMPT = ['contrarian', 'contrarian_enhanced', 'enhanced_mr', 'bb_bounce']

Lesson: If you slap a trend filter on a contrarian strategy, you kill the strategy’s entire reason for existing. Don’t tell a bot “never buy during drops” when drops are the whole point.

Lessons Learned

“Buy when fearful” depends on your time horizon. On a yearly scale, it works. On an hourly scale, it can keep falling.
Extreme Fear DCA = losing more in a downtrend. Your average entry gets better, but your position gets bigger — and so do your losses.
FGI alone isn’t enough. You need crash depth, RSI, and volume to judge whether you’re actually near a real bottom.
Tier-based entries beat DCA. Scaling by depth of fear (tiers) instead of time (DCA) prevents the full-size-into-shallow-dip mistake.
Never put a trend filter on a contrarian strategy. If drops are opportunities, “block entries during drops” kills the strategy.
Patience is the strategy. FGI ≤ 10 happens once or twice a year. That’s what the other 28 bots are for.

Previous post: Regime System — When the Market Has Four Faces