• Ian Domowitz

  • Kumar Giritharan


Social Media Sentiment and the FX Market


This is a study of the effect of social media sentiment on pricing and trading fundamentals in the foreign exchange market. Examination of the predictability of prices, the cost and depth of liquidity and the topography of the global limit order book for a currency pair is the focus. Stand-alone forecasting exercises are combined with event studies in order to reach conclusions. There is some promise in price forecasting results, but work remains to be done in terms of strategy and hedging applications. Results with respect to FX market structure are less encouraging, with the possible exception of a tight correlation between sentiment and the cost of liquidity.

The use of “alternative data” in financial markets is on the rise, and social media information is the star of this trend. The use of social sentiment information is largely confined to stock market prices, including overall market reactions, single stock movements and ETF returns.i

We introduce the use of social media data in the analysis of foreign exchange movements. Our interest is broader than exchange rate pricing, extending to short-run changes in quantifiable FX market structure. Focus centers on the following questions:

  • What combinations of social media sentiment information may provide predictive value with respect to exchange rate pricing?
  • Does sentiment information lead changes in the cost of liquidity for larger deal sizes?
  • Is there information in sentiment data, which forecasts changes in the depth of the global limit order book, as well as associated changes in volatility and trader interaction in the market?

The first question is of interest to those designing hedging strategies, as well as those who trade foreign exchange for profit or provide forecasts to such traders. The next two queries are oriented toward institutional investors, who trade currencies in the normal course of business, offsetting currency positions in the rebalancing of global securities portfolios. Previous work shows, for example, that a large institution incurs between $14 million and $40 million in equity-linked FX trading costs annually, depending on trading strategy.ii The cost of liquidity may be calculated from tradable quotes and, based on a variety of data streams, a limit order book can be constructed that provides measures of liquidity for each potential price point.

Predictive power for pricing at within-day intervals and across days is measured by the probability of directional correctness; i.e., does an increase/decrease in social media sentiment correlate with an increase/decrease in price in the near future. Directional correctness can be as high as 60% in several cases, rising to 70% for a strategy, which exploits both days of a two-day window. Two lessons are learned in the process. First, the FX market cannot be aggregated with any success; results are perforce currency pair by currency pair. This differs sharply from results on sentiment and equity investing. Second, the degree of social media traffic matters. This is consistent with the notion that social sentiment reflects the wisdom of crowds. The bigger the crowd, the better the
result, in principle. No direct relationship between sentiment and FX volatility can be isolated, however.

The price prediction results extend into two case studies. The first is an example of an unanticipated event in the equities market and its effect on a non-local currency. The second is more classic, a somewhat anticipated central bank announcement, and its effect on the local currency.

The case studies extend the scope of analysis into quantifiable short-term FX market structure. Sentiment has little effect on the standard measure of the bid-ask spread. On the other hand, the cost of liquidity for large deal sizes is tightly correlated with sentiment, although a direct connection is difficult to isolate. We find no effect of sentiment on the number of active traders in the market, and there are ways to explain changes in depth of the global limit order book for a currency pair that do not depend on sentiment.

These conclusions are reached based on the combination of two alternative data sets, over the period January 2014 through May 2016. Social media sentiment information is sourced from Social Market Analytics.iii Fifteen-minute updates are derived from both Twitter and StockTwits message streams. Twitter is well known, with 310 million active users as of the middle of 2016. StockTwits organizes streams of information around currency pairs, among other financial instruments. There are roughly 300,000 market professionals active on this network. StockTwits focuses solely on investing and filters out unrelated messages. For our purpose here, Social Market Analytics extracts data to capture only tweets containing commentary on currency pairs, and analyzes each message for financial market relevance. Natural language processing algorithms are employed to determine sentiment levels by currency pair, which are then normalized as a Z-score over a fixed look-back interval. These data then are enriched with pricing and volatility information for five major currency pairs, AUDUSD, EURUSD, GBPUSD, USDCAD and USDJPY.

The second data set is our own derived data, based on tradable dealer quotes from six global banks and five major ECNs. Described in detail elsewhere, a global limit order book is constructed that delivers size-adjusted spreads for order sizes up to USD 200 million.iv From these spreads, the cost of liquidity may be calculated for relevant deal sizes. The data also permit the evaluation of the depth of the order book at various price levels, as well as the extent of trader participation in the market for any given currency pair. Information on levels of volatility is supplemented by empirical distributions of volatility, conditioned on pair, DST regime, time of day and quote consolidation type.v

Price prediction

From an FX pricing perspective, sentiment information is no different from any other technical trading data. Said differently, it presents all the typical data mining issues. We make no attempt to dig deeply, however. Rather, interest centers on illustrating directions and pitfalls in the exercise. All results are based on out-of-sample forecasts of rate direction and are characterized by the probability of directional correctness.

The first exercise, illustrated in Exhibits 1 and 2, does not depend on data mining. Rather, the parameters of the forecasting problem are fixed in advance. Forecasts are over a four-hour window, with a look-back period of four hours; the forecasts are updated every 15 minutes. Tests of predictability are limited to look-back windows for which the currency pair exhibits virtually no price movement. Prior trends induce social sentiment following, and we are looking for new information, which is not reflected in trending prices. Positive and negative sentiment scores are grouped in three categories, based on the size of the sentiment change. The level of social activity, measured by number of tweets, is similarly aggregated. The metric of success is simple: the percentage of times that a positive/negative sentiment change leads to a positive/negative change in price over the four-hour forecast window. 

Read the full report >